@elizaos/plugin-memory 1.0.4 → 1.0.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -14,21 +14,29 @@ Advanced memory management plugin for ElizaOS that provides intelligent conversa
14
14
  ### 🧠 Long-term Memory (Persistent Facts)
15
15
 
16
16
  - **Intelligent Extraction**: Automatically learns facts about users from conversations
17
- - **Categorized Storage**: Organizes information into 9 semantic categories
17
+ - **Cognitive Science Based**: Organizes information into 3 core memory types (episodic, semantic, procedural)
18
+ - **Strict Criteria**: Only extracts truly significant, persistent information
18
19
  - **Confidence Scoring**: Tracks reliability of stored information
19
- - **Cross-session Persistence**: Remembers user preferences and context across all interactions
20
+ - **Cross-session Persistence**: Remembers user context across all interactions
20
21
 
21
- ### 📊 Memory Categories
22
+ ### 📊 Memory Categories (Based on Cognitive Science)
22
23
 
23
- 1. **Identity**: User's name, role, profession (e.g., "I'm a data scientist")
24
- 2. **Expertise**: Domain knowledge, skills, familiarity with topics
25
- 3. **Projects**: Ongoing work, past interactions, recurring topics
26
- 4. **Preferences**: Communication style, format preferences, verbosity
27
- 5. **Data Sources**: Frequently used files, databases, APIs
28
- 6. **Goals**: Broader intentions and objectives
29
- 7. **Constraints**: User-defined rules or limitations
30
- 8. **Definitions**: Custom terms, acronyms, glossaries
31
- 9. **Behavioral Patterns**: Interaction styles and tendencies
24
+ The plugin uses the three fundamental types of long-term memory from cognitive science:
25
+
26
+ 1. **Episodic Memory**: Personal experiences and specific events
27
+ - Example: "User completed migration from MongoDB to PostgreSQL in Q2 2024"
28
+ - Contains: WHO did WHAT, WHEN/WHERE
29
+ - Use for: Significant project milestones, important incidents, formative experiences
30
+
31
+ 2. **Semantic Memory**: General facts, concepts, and knowledge
32
+ - Example: "User is a senior TypeScript developer with 8 years experience"
33
+ - Contains: Factual, timeless information
34
+ - Use for: Professional identity, core expertise, established facts about work context
35
+
36
+ 3. **Procedural Memory**: Skills, workflows, and how-to knowledge
37
+ - Example: "User follows TDD workflow: writes tests first, then implementation"
38
+ - Contains: HOW user does things
39
+ - Use for: Consistent workflows, methodologies, debugging processes
32
40
 
33
41
  ## Installation
34
42
 
@@ -58,12 +66,16 @@ Configure the plugin via environment variables in your `.env` file:
58
66
 
59
67
  ```env
60
68
  # Short-term Memory Settings
61
- MEMORY_SUMMARIZATION_THRESHOLD=50 # Messages before summarization (default: 50)
62
- MEMORY_RETAIN_RECENT=10 # Recent messages to keep (default: 10)
69
+ MEMORY_SUMMARIZATION_THRESHOLD=16 # Messages before summarization starts (default: 16)
70
+ MEMORY_SUMMARIZATION_INTERVAL=10 # Update summary every N messages (default: 10)
71
+ MEMORY_RETAIN_RECENT=10 # Recent messages to keep (default: 10)
72
+ MEMORY_MAX_NEW_MESSAGES=20 # Max new messages in summary update (default: 20)
63
73
 
64
74
  # Long-term Memory Settings
65
- MEMORY_LONG_TERM_ENABLED=true # Enable long-term extraction (default: true)
66
- MEMORY_CONFIDENCE_THRESHOLD=0.7 # Minimum confidence to store (default: 0.7)
75
+ MEMORY_LONG_TERM_ENABLED=true # Enable long-term extraction (default: true)
76
+ MEMORY_EXTRACTION_THRESHOLD=30 # Min messages before extraction starts (default: 30)
77
+ MEMORY_EXTRACTION_INTERVAL=10 # Run extraction every N messages (default: 10)
78
+ MEMORY_CONFIDENCE_THRESHOLD=0.85 # Minimum confidence to store (default: 0.85)
67
79
  ```
68
80
 
69
81
  ### Manual Memory Storage
@@ -72,13 +84,13 @@ Users can explicitly ask the agent to remember information:
72
84
 
73
85
  ```
74
86
  User: "Remember that I prefer TypeScript over JavaScript"
75
- Agent: I've made a note of that in my Preferences memory: "User prefers TypeScript over JavaScript"
87
+ Agent: I've made a note of that in my Semantic memory: "User prefers TypeScript over JavaScript"
76
88
 
77
89
  User: "Keep in mind I'm working on a startup project"
78
- Agent: I've made a note of that in my Projects memory: "User is working on a startup project"
90
+ Agent: I've made a note of that in my Episodic memory: "User is working on a startup project"
79
91
 
80
- User: "Don't forget I use Python 3.11"
81
- Agent: I've made a note of that in my Data Sources memory: "User uses Python 3.11"
92
+ User: "Don't forget I always use TDD"
93
+ Agent: I've made a note of that in my Procedural memory: "User follows TDD (Test-Driven Development) methodology"
82
94
  ```
83
95
 
84
96
  ### Accessing the Memory Service
@@ -93,7 +105,7 @@ const memoryService = runtime.getService('memory') as MemoryService;
93
105
  await memoryService.storeLongTermMemory({
94
106
  agentId: runtime.agentId,
95
107
  entityId: userId,
96
- category: LongTermMemoryCategory.PREFERENCES,
108
+ category: LongTermMemoryCategory.SEMANTIC,
97
109
  content: 'User prefers concise responses',
98
110
  confidence: 0.9,
99
111
  source: 'manual',
@@ -170,11 +182,29 @@ No manual migration is required - the schema is handled automatically by the run
170
182
 
171
183
  ### Long-term Memory Flow
172
184
 
173
- 1. **Monitoring**: longTermExtractionEvaluator runs periodically (every 10 messages)
174
- 2. **Analysis**: LLM analyzes conversation for facts worth remembering
175
- 3. **Extraction**: Identifies facts, categorizes them, assigns confidence
176
- 4. **Storage**: High-confidence facts stored in long_term_memories table
177
- 5. **Retrieval**: longTermMemoryProvider injects relevant facts in all future conversations
185
+ 1. **Warm-up Period**: Extraction waits until 30+ messages (configurable) to ensure meaningful patterns
186
+ 2. **Monitoring**: longTermExtractionEvaluator runs periodically (every 10 messages after threshold)
187
+ 3. **Analysis**: LLM analyzes conversation for **persistent, important** facts worth remembering
188
+ 4. **Strict Filtering**: Applies cognitive science principles to extract only truly significant information
189
+ 5. **Storage**: High-confidence facts (≥0.85) stored in long_term_memories table
190
+ 6. **Retrieval**: longTermMemoryProvider injects relevant facts in all future conversations
191
+
192
+ **Ultra-Strict Extraction Criteria**: The evaluator uses stringent criteria to prevent memory pollution:
193
+
194
+ - ✅ **DO Extract:**
195
+ - **Episodic**: Significant milestones, important incidents, major decisions with lasting impact
196
+ - **Semantic**: Professional identity, core expertise, established facts (explicitly stated or conclusively demonstrated)
197
+ - **Procedural**: Consistent workflows (3+ occurrences or explicitly stated), standard practices, methodologies
198
+
199
+ - ❌ **NEVER Extract:**
200
+ - One-time requests or tasks
201
+ - Casual conversations without lasting significance
202
+ - Exploratory questions or testing
203
+ - Temporary context or situational information
204
+ - Preferences from single occurrence
205
+ - Social pleasantries
206
+ - Common patterns everyone has
207
+ - General knowledge not specific to user
178
208
 
179
209
  ### Manual Memory Flow
180
210
 
@@ -221,10 +251,19 @@ No manual migration is required - the schema is handled automatically by the run
221
251
 
222
252
  ### Configuration Tips
223
253
 
224
- - **High-frequency chatbots**: Lower threshold (30-40 messages)
225
- - **Long-form conversations**: Higher threshold (60-100 messages)
226
- - **Critical applications**: Higher confidence threshold (0.8-0.9)
227
- - **Exploratory use**: Lower confidence threshold (0.6-0.7)
254
+ **Short-term Memory:**
255
+ - **High-frequency chatbots**: Lower summarization threshold (10-15 messages)
256
+ - **Long-form conversations**: Higher threshold (20-30 messages)
257
+ - **Adjust retention**: Keep more recent messages for immediate context
258
+
259
+ **Long-term Memory:**
260
+ - **Conservative extraction**: Keep threshold at 30+ messages for better pattern recognition (default)
261
+ - **Aggressive extraction**: Lower threshold to 20 messages if needed (may reduce quality)
262
+ - **Balanced approach**: Default 0.85 confidence threshold ensures high-quality extractions
263
+ - **More permissive**: Lower confidence to 0.80 for more extractions (risk of lower quality)
264
+ - **Most strict**: Raise confidence to 0.90 for only the most certain facts
265
+ - **Frequent updates**: Lower extraction interval to 5-8 messages for faster learning
266
+ - **Conservative updates**: Keep default 10+ message interval to prevent over-extraction
228
267
 
229
268
  ## Advanced Features
230
269
 
@@ -246,7 +285,7 @@ Use the `memory_access_logs` table to:
246
285
 
247
286
  ### Custom Categories
248
287
 
249
- Extend `LongTermMemoryCategory` enum for domain-specific categories:
288
+ The plugin uses three scientifically-grounded memory types from cognitive science. If you need additional categories for domain-specific use cases, you can extend the enum:
250
289
 
251
290
  ```typescript
252
291
  export enum CustomMemoryCategory {
@@ -256,6 +295,8 @@ export enum CustomMemoryCategory {
256
295
  }
257
296
  ```
258
297
 
298
+ **Note**: Consider carefully whether your custom category truly represents a different type of memory, or if it can be classified under episodic (events), semantic (facts), or procedural (how-to) memory.
299
+
259
300
  ## Testing
260
301
 
261
302
  Run the test suite:
@@ -1,8 +1,8 @@
1
- var zJ=Object.defineProperty;var DJ=(J,Q)=>{for(var Y in Q)zJ(J,Y,{get:Q[Y],enumerable:!0,configurable:!0,set:(W)=>Q[Y]=()=>W})};import{Service as MJ,logger as H}from"@elizaos/core";import{eq as D,and as g,desc as p,sql as jJ,cosineDistance as yJ,gte as lJ}from"drizzle-orm";var JJ={};DJ(JJ,{sessionSummaries:()=>q,memoryAccessLogs:()=>XJ,longTermMemories:()=>X});import{sql as $J}from"drizzle-orm";import{pgTable as RJ,text as m,integer as bJ,jsonb as CJ,real as AJ,index as l,varchar as s,timestamp as t}from"drizzle-orm/pg-core";var X=RJ("long_term_memories",{id:s("id",{length:36}).primaryKey(),agentId:s("agent_id",{length:36}).notNull(),entityId:s("entity_id",{length:36}).notNull(),category:m("category").notNull(),content:m("content").notNull(),metadata:CJ("metadata"),embedding:AJ("embedding").array(),confidence:AJ("confidence").default(1),source:m("source"),createdAt:t("created_at").default($J`now()`).notNull(),updatedAt:t("updated_at").default($J`now()`).notNull(),lastAccessedAt:t("last_accessed_at"),accessCount:bJ("access_count").default(0)},(J)=>({agentEntityIdx:l("long_term_memories_agent_entity_idx").on(J.agentId,J.entityId),categoryIdx:l("long_term_memories_category_idx").on(J.category),confidenceIdx:l("long_term_memories_confidence_idx").on(J.confidence),createdAtIdx:l("long_term_memories_created_at_idx").on(J.createdAt)}));import{sql as UJ}from"drizzle-orm";import{pgTable as fJ,text as hJ,integer as VJ,jsonb as GJ,real as TJ,index as r,varchar as c,timestamp as d}from"drizzle-orm/pg-core";var q=fJ("session_summaries",{id:c("id",{length:36}).primaryKey(),agentId:c("agent_id",{length:36}).notNull(),roomId:c("room_id",{length:36}).notNull(),entityId:c("entity_id",{length:36}),summary:hJ("summary").notNull(),messageCount:VJ("message_count").notNull(),lastMessageOffset:VJ("last_message_offset").notNull().default(0),startTime:d("start_time").notNull(),endTime:d("end_time").notNull(),topics:GJ("topics"),metadata:GJ("metadata"),embedding:TJ("embedding").array(),createdAt:d("created_at").default(UJ`now()`).notNull(),updatedAt:d("updated_at").default(UJ`now()`).notNull()},(J)=>({agentRoomIdx:r("session_summaries_agent_room_idx").on(J.agentId,J.roomId),entityIdx:r("session_summaries_entity_idx").on(J.entityId),startTimeIdx:r("session_summaries_start_time_idx").on(J.startTime)}));import{sql as vJ}from"drizzle-orm";import{pgTable as IJ,text as SJ,integer as xJ,real as pJ,index as e,varchar as u,timestamp as wJ}from"drizzle-orm/pg-core";var XJ=IJ("memory_access_logs",{id:u("id",{length:36}).primaryKey(),agentId:u("agent_id",{length:36}).notNull(),memoryId:u("memory_id",{length:36}).notNull(),memoryType:SJ("memory_type").notNull(),accessedAt:wJ("accessed_at").default(vJ`now()`).notNull(),roomId:u("room_id",{length:36}),relevanceScore:pJ("relevance_score"),wasUseful:xJ("was_useful")},(J)=>({memoryIdx:e("memory_access_logs_memory_idx").on(J.memoryId),agentIdx:e("memory_access_logs_agent_idx").on(J.agentId),accessedAtIdx:e("memory_access_logs_accessed_at_idx").on(J.accessedAt)}));class w extends MJ{static serviceType="memory";sessionMessageCounts;memoryConfig;lastExtractionCheckpoints;capabilityDescription="Advanced memory management with short-term summarization and long-term persistent facts";constructor(J){super(J);this.sessionMessageCounts=new Map,this.lastExtractionCheckpoints=new Map,this.memoryConfig={shortTermSummarizationThreshold:5,shortTermRetainRecent:10,longTermExtractionEnabled:!0,longTermVectorSearchEnabled:!1,longTermConfidenceThreshold:0.7,longTermExtractionInterval:5,summaryModelType:"TEXT_LARGE",summaryMaxTokens:2500}}static async start(J){let Q=new w(J);return await Q.initialize(J),Q}async stop(){H.info("MemoryService stopped")}async initialize(J){this.runtime=J;let Q=J.getSetting("MEMORY_SUMMARIZATION_THRESHOLD");if(Q)this.memoryConfig.shortTermSummarizationThreshold=parseInt(Q,10);let Y=J.getSetting("MEMORY_RETAIN_RECENT");if(Y)this.memoryConfig.shortTermRetainRecent=parseInt(Y,10);let W=J.getSetting("MEMORY_LONG_TERM_ENABLED");if(W==="false")this.memoryConfig.longTermExtractionEnabled=!1;else if(W==="true")this.memoryConfig.longTermExtractionEnabled=!0;let K=J.getSetting("MEMORY_CONFIDENCE_THRESHOLD");if(K)this.memoryConfig.longTermConfidenceThreshold=parseFloat(K);H.info({summarizationThreshold:this.memoryConfig.shortTermSummarizationThreshold,retainRecent:this.memoryConfig.shortTermRetainRecent,longTermEnabled:this.memoryConfig.longTermExtractionEnabled,extractionInterval:this.memoryConfig.longTermExtractionInterval,confidenceThreshold:this.memoryConfig.longTermConfidenceThreshold},"MemoryService initialized")}getDb(){let J=this.runtime.db;if(!J)throw Error("Database not available");return J}getConfig(){return{...this.memoryConfig}}updateConfig(J){this.memoryConfig={...this.memoryConfig,...J}}incrementMessageCount(J){let Y=(this.sessionMessageCounts.get(J)||0)+1;return this.sessionMessageCounts.set(J,Y),Y}resetMessageCount(J){this.sessionMessageCounts.set(J,0)}async shouldSummarize(J){return await this.runtime.countMemories(J,!1,"messages")>=this.memoryConfig.shortTermSummarizationThreshold}getExtractionKey(J,Q){return`memory:extraction:${J}:${Q}`}async getLastExtractionCheckpoint(J,Q){let Y=this.getExtractionKey(J,Q),W=this.lastExtractionCheckpoints.get(Y);if(W!==void 0)return W;try{let Z=await this.runtime.getCache(Y)??0;return this.lastExtractionCheckpoints.set(Y,Z),Z}catch(K){return H.warn({error:K},"Failed to get extraction checkpoint from cache"),0}}async setLastExtractionCheckpoint(J,Q,Y){let W=this.getExtractionKey(J,Q);this.lastExtractionCheckpoints.set(W,Y);try{await this.runtime.setCache(W,Y),H.debug(`Set extraction checkpoint for ${J} in room ${Q} at message count ${Y}`)}catch(K){H.error({error:K},"Failed to persist extraction checkpoint to cache")}}async shouldRunExtraction(J,Q,Y){let W=this.memoryConfig.longTermExtractionInterval,K=await this.getLastExtractionCheckpoint(J,Q),Z=Math.floor(Y/W)*W,$=Y>=W&&Z>K;return H.debug({entityId:J,roomId:Q,currentMessageCount:Y,interval:W,lastCheckpoint:K,currentCheckpoint:Z,shouldRun:$},"Extraction check"),$}async storeLongTermMemory(J){let Q=this.getDb(),Y=crypto.randomUUID(),W=new Date,K={id:Y,createdAt:W,updatedAt:W,accessCount:0,...J};try{await Q.insert(X).values({id:K.id,agentId:K.agentId,entityId:K.entityId,category:K.category,content:K.content,metadata:K.metadata||{},embedding:K.embedding,confidence:K.confidence,source:K.source,accessCount:K.accessCount,createdAt:W,updatedAt:W,lastAccessedAt:K.lastAccessedAt})}catch(Z){throw H.error({error:Z},"Failed to store long-term memory"),Z}return H.info(`Stored long-term memory: ${K.category} for entity ${K.entityId}`),K}async getLongTermMemories(J,Q,Y=10){let W=this.getDb(),K=[D(X.agentId,this.runtime.agentId),D(X.entityId,J)];if(Q)K.push(D(X.category,Q));return(await W.select().from(X).where(g(...K)).orderBy(p(X.confidence),p(X.updatedAt)).limit(Y)).map(($)=>({id:$.id,agentId:$.agentId,entityId:$.entityId,category:$.category,content:$.content,metadata:$.metadata,embedding:$.embedding,confidence:$.confidence,source:$.source,createdAt:$.createdAt,updatedAt:$.updatedAt,lastAccessedAt:$.lastAccessedAt,accessCount:$.accessCount}))}async updateLongTermMemory(J,Q){let Y=this.getDb(),W={updatedAt:new Date};if(Q.content!==void 0)W.content=Q.content;if(Q.metadata!==void 0)W.metadata=Q.metadata;if(Q.confidence!==void 0)W.confidence=Q.confidence;if(Q.embedding!==void 0)W.embedding=Q.embedding;if(Q.lastAccessedAt!==void 0)W.lastAccessedAt=Q.lastAccessedAt;if(Q.accessCount!==void 0)W.accessCount=Q.accessCount;await Y.update(X).set(W).where(D(X.id,J)),H.info(`Updated long-term memory: ${J}`)}async deleteLongTermMemory(J){await this.getDb().delete(X).where(D(X.id,J)),H.info(`Deleted long-term memory: ${J}`)}async getCurrentSessionSummary(J){let Y=await this.getDb().select().from(q).where(g(D(q.agentId,this.runtime.agentId),D(q.roomId,J))).orderBy(p(q.updatedAt)).limit(1);if(Y.length===0)return null;let W=Y[0];return{id:W.id,agentId:W.agentId,roomId:W.roomId,entityId:W.entityId,summary:W.summary,messageCount:W.messageCount,lastMessageOffset:W.lastMessageOffset,startTime:W.startTime,endTime:W.endTime,topics:W.topics||[],metadata:W.metadata,embedding:W.embedding,createdAt:W.createdAt,updatedAt:W.updatedAt}}async storeSessionSummary(J){let Q=this.getDb(),Y=crypto.randomUUID(),W=new Date,K={id:Y,createdAt:W,updatedAt:W,...J};return await Q.insert(q).values({id:K.id,agentId:K.agentId,roomId:K.roomId,entityId:K.entityId||null,summary:K.summary,messageCount:K.messageCount,lastMessageOffset:K.lastMessageOffset,startTime:K.startTime,endTime:K.endTime,topics:K.topics||[],metadata:K.metadata||{},embedding:K.embedding,createdAt:W,updatedAt:W}),H.info(`Stored session summary for room ${K.roomId}`),K}async updateSessionSummary(J,Q){let Y=this.getDb(),W={updatedAt:new Date};if(Q.summary!==void 0)W.summary=Q.summary;if(Q.messageCount!==void 0)W.messageCount=Q.messageCount;if(Q.lastMessageOffset!==void 0)W.lastMessageOffset=Q.lastMessageOffset;if(Q.endTime!==void 0)W.endTime=Q.endTime;if(Q.topics!==void 0)W.topics=Q.topics;if(Q.metadata!==void 0)W.metadata=Q.metadata;if(Q.embedding!==void 0)W.embedding=Q.embedding;await Y.update(q).set(W).where(D(q.id,J)),H.info(`Updated session summary: ${J}`)}async getSessionSummaries(J,Q=5){return(await this.getDb().select().from(q).where(g(D(q.agentId,this.runtime.agentId),D(q.roomId,J))).orderBy(p(q.updatedAt)).limit(Q)).map((K)=>({id:K.id,agentId:K.agentId,roomId:K.roomId,entityId:K.entityId,summary:K.summary,messageCount:K.messageCount,lastMessageOffset:K.lastMessageOffset,startTime:K.startTime,endTime:K.endTime,topics:K.topics||[],metadata:K.metadata,embedding:K.embedding,createdAt:K.createdAt,updatedAt:K.updatedAt}))}async searchLongTermMemories(J,Q,Y=5,W=0.7){if(!this.memoryConfig.longTermVectorSearchEnabled)return H.warn("Vector search is not enabled, falling back to recent memories"),this.getLongTermMemories(J,void 0,Y);let K=this.getDb();try{let Z=Q.map((_)=>Number.isFinite(_)?Number(_.toFixed(6)):0),$=jJ`1 - (${yJ(X.embedding,Z)})`,U=[D(X.agentId,this.runtime.agentId),D(X.entityId,J),jJ`${X.embedding} IS NOT NULL`];if(W>0)U.push(lJ($,W));return(await K.select({memory:X,similarity:$}).from(X).where(g(...U)).orderBy(p($)).limit(Y)).map((_)=>({id:_.memory.id,agentId:_.memory.agentId,entityId:_.memory.entityId,category:_.memory.category,content:_.memory.content,metadata:_.memory.metadata,embedding:_.memory.embedding,confidence:_.memory.confidence,source:_.memory.source,createdAt:_.memory.createdAt,updatedAt:_.memory.updatedAt,lastAccessedAt:_.memory.lastAccessedAt,accessCount:_.memory.accessCount,similarity:_.similarity}))}catch(Z){return H.warn({error:Z},"Vector search failed, falling back to recent memories"),this.getLongTermMemories(J,void 0,Y)}}async getFormattedLongTermMemories(J){let Q=await this.getLongTermMemories(J,void 0,20);if(Q.length===0)return"";let Y=new Map;for(let K of Q){if(!Y.has(K.category))Y.set(K.category,[]);Y.get(K.category)?.push(K)}let W=[];for(let[K,Z]of Y.entries()){let $=K.split("_").map((V)=>V.charAt(0).toUpperCase()+V.slice(1)).join(" "),U=Z.map((V)=>`- ${V.content}`).join(`
2
- `);W.push(`**${$}**:
3
- ${U}`)}return W.join(`
1
+ var MJ=Object.defineProperty;var wJ=(J,Y)=>{for(var Q in Y)MJ(J,Q,{get:Y[Q],enumerable:!0,configurable:!0,set:(Z)=>Y[Q]=()=>Z})};import{Service as rJ,logger as f}from"@elizaos/core";import{eq as k,and as i,desc as JJ,sql as CJ,cosineDistance as tJ,gte as eJ}from"drizzle-orm";var kJ={};wJ(kJ,{sessionSummaries:()=>D,memoryAccessLogs:()=>bJ,longTermMemories:()=>H});import{sql as zJ}from"drizzle-orm";import{pgTable as yJ,text as OJ,integer as lJ,jsonb as cJ,real as qJ,index as VJ,varchar as HJ,timestamp as PJ}from"drizzle-orm/pg-core";var H=yJ("long_term_memories",{id:HJ("id",{length:36}).primaryKey(),agentId:HJ("agent_id",{length:36}).notNull(),entityId:HJ("entity_id",{length:36}).notNull(),category:OJ("category").notNull(),content:OJ("content").notNull(),metadata:cJ("metadata"),embedding:qJ("embedding").array(),confidence:qJ("confidence").default(1),source:OJ("source"),createdAt:PJ("created_at").default(zJ`now()`).notNull(),updatedAt:PJ("updated_at").default(zJ`now()`).notNull(),lastAccessedAt:PJ("last_accessed_at"),accessCount:lJ("access_count").default(0)},(J)=>({agentEntityIdx:VJ("long_term_memories_agent_entity_idx").on(J.agentId,J.entityId),categoryIdx:VJ("long_term_memories_category_idx").on(J.category),confidenceIdx:VJ("long_term_memories_confidence_idx").on(J.confidence),createdAtIdx:VJ("long_term_memories_created_at_idx").on(J.createdAt)}));import{sql as DJ}from"drizzle-orm";import{pgTable as dJ,text as uJ,integer as RJ,jsonb as fJ,real as gJ,index as LJ,varchar as WJ,timestamp as AJ}from"drizzle-orm/pg-core";var D=dJ("session_summaries",{id:WJ("id",{length:36}).primaryKey(),agentId:WJ("agent_id",{length:36}).notNull(),roomId:WJ("room_id",{length:36}).notNull(),entityId:WJ("entity_id",{length:36}),summary:uJ("summary").notNull(),messageCount:RJ("message_count").notNull(),lastMessageOffset:RJ("last_message_offset").notNull().default(0),startTime:AJ("start_time").notNull(),endTime:AJ("end_time").notNull(),topics:fJ("topics"),metadata:fJ("metadata"),embedding:gJ("embedding").array(),createdAt:AJ("created_at").default(DJ`now()`).notNull(),updatedAt:AJ("updated_at").default(DJ`now()`).notNull()},(J)=>({agentRoomIdx:LJ("session_summaries_agent_room_idx").on(J.agentId,J.roomId),entityIdx:LJ("session_summaries_entity_idx").on(J.entityId),startTimeIdx:LJ("session_summaries_start_time_idx").on(J.startTime)}));import{sql as iJ}from"drizzle-orm";import{pgTable as nJ,text as aJ,integer as mJ,real as oJ,index as NJ,varchar as GJ,timestamp as sJ}from"drizzle-orm/pg-core";var bJ=nJ("memory_access_logs",{id:GJ("id",{length:36}).primaryKey(),agentId:GJ("agent_id",{length:36}).notNull(),memoryId:GJ("memory_id",{length:36}).notNull(),memoryType:aJ("memory_type").notNull(),accessedAt:sJ("accessed_at").default(iJ`now()`).notNull(),roomId:GJ("room_id",{length:36}),relevanceScore:oJ("relevance_score"),wasUseful:mJ("was_useful")},(J)=>({memoryIdx:NJ("memory_access_logs_memory_idx").on(J.memoryId),agentIdx:NJ("memory_access_logs_agent_idx").on(J.agentId),accessedAtIdx:NJ("memory_access_logs_accessed_at_idx").on(J.accessedAt)}));class KJ extends rJ{static serviceType="memory";sessionMessageCounts;memoryConfig;lastExtractionCheckpoints;capabilityDescription="Advanced memory management with short-term summarization and long-term persistent facts";constructor(J){super(J);this.sessionMessageCounts=new Map,this.lastExtractionCheckpoints=new Map,this.memoryConfig={shortTermSummarizationThreshold:16,shortTermRetainRecent:6,shortTermSummarizationInterval:10,longTermExtractionEnabled:!0,longTermVectorSearchEnabled:!1,longTermConfidenceThreshold:0.85,longTermExtractionThreshold:30,longTermExtractionInterval:10,summaryModelType:"TEXT_LARGE",summaryMaxTokens:2500,summaryMaxNewMessages:20}}static async start(J){let Y=new KJ(J);return await Y.initialize(J),Y}async stop(){f.info("MemoryService stopped")}async initialize(J){this.runtime=J;let Y=J.getSetting("MEMORY_SUMMARIZATION_THRESHOLD");if(Y)this.memoryConfig.shortTermSummarizationThreshold=parseInt(Y,10);let Q=J.getSetting("MEMORY_RETAIN_RECENT");if(Q)this.memoryConfig.shortTermRetainRecent=parseInt(Q,10);let Z=J.getSetting("MEMORY_SUMMARIZATION_INTERVAL");if(Z)this.memoryConfig.shortTermSummarizationInterval=parseInt(Z,10);let K=J.getSetting("MEMORY_MAX_NEW_MESSAGES");if(K)this.memoryConfig.summaryMaxNewMessages=parseInt(K,10);let _=J.getSetting("MEMORY_LONG_TERM_ENABLED");if(_==="false")this.memoryConfig.longTermExtractionEnabled=!1;else if(_==="true")this.memoryConfig.longTermExtractionEnabled=!0;let $=J.getSetting("MEMORY_CONFIDENCE_THRESHOLD");if($)this.memoryConfig.longTermConfidenceThreshold=parseFloat($);let U=J.getSetting("MEMORY_EXTRACTION_THRESHOLD");if(U)this.memoryConfig.longTermExtractionThreshold=parseInt(U,10);let P=J.getSetting("MEMORY_EXTRACTION_INTERVAL");if(P)this.memoryConfig.longTermExtractionInterval=parseInt(P,10);f.info({summarizationThreshold:this.memoryConfig.shortTermSummarizationThreshold,summarizationInterval:this.memoryConfig.shortTermSummarizationInterval,maxNewMessages:this.memoryConfig.summaryMaxNewMessages,retainRecent:this.memoryConfig.shortTermRetainRecent,longTermEnabled:this.memoryConfig.longTermExtractionEnabled,extractionThreshold:this.memoryConfig.longTermExtractionThreshold,extractionInterval:this.memoryConfig.longTermExtractionInterval,confidenceThreshold:this.memoryConfig.longTermConfidenceThreshold},"MemoryService initialized")}getDb(){let J=this.runtime.db;if(!J)throw Error("Database not available");return J}getConfig(){return{...this.memoryConfig}}updateConfig(J){this.memoryConfig={...this.memoryConfig,...J}}incrementMessageCount(J){let Q=(this.sessionMessageCounts.get(J)||0)+1;return this.sessionMessageCounts.set(J,Q),Q}resetMessageCount(J){this.sessionMessageCounts.set(J,0)}async shouldSummarize(J){return await this.runtime.countMemories(J,!1,"messages")>=this.memoryConfig.shortTermSummarizationThreshold}getExtractionKey(J,Y){return`memory:extraction:${J}:${Y}`}async getLastExtractionCheckpoint(J,Y){let Q=this.getExtractionKey(J,Y),Z=this.lastExtractionCheckpoints.get(Q);if(Z!==void 0)return Z;try{let _=await this.runtime.getCache(Q)??0;return this.lastExtractionCheckpoints.set(Q,_),_}catch(K){return f.warn({error:K},"Failed to get extraction checkpoint from cache"),0}}async setLastExtractionCheckpoint(J,Y,Q){let Z=this.getExtractionKey(J,Y);this.lastExtractionCheckpoints.set(Z,Q);try{await this.runtime.setCache(Z,Q),f.debug(`Set extraction checkpoint for ${J} in room ${Y} at message count ${Q}`)}catch(K){f.error({error:K},"Failed to persist extraction checkpoint to cache")}}async shouldRunExtraction(J,Y,Q){let Z=this.memoryConfig.longTermExtractionThreshold,K=this.memoryConfig.longTermExtractionInterval;if(Q<Z)return f.debug({entityId:J,roomId:Y,currentMessageCount:Q,threshold:Z,shouldRun:!1},"Extraction check: below threshold"),!1;let _=await this.getLastExtractionCheckpoint(J,Y),$=Math.floor(Q/K)*K,U=Q>=Z&&$>_;return f.debug({entityId:J,roomId:Y,currentMessageCount:Q,threshold:Z,interval:K,lastCheckpoint:_,currentCheckpoint:$,shouldRun:U},"Extraction check"),U}async storeLongTermMemory(J){let Y=this.getDb(),Q=crypto.randomUUID(),Z=new Date,K={id:Q,createdAt:Z,updatedAt:Z,accessCount:0,...J};try{await Y.insert(H).values({id:K.id,agentId:K.agentId,entityId:K.entityId,category:K.category,content:K.content,metadata:K.metadata||{},embedding:K.embedding,confidence:K.confidence,source:K.source,accessCount:K.accessCount,createdAt:Z,updatedAt:Z,lastAccessedAt:K.lastAccessedAt})}catch(_){throw f.error({error:_},"Failed to store long-term memory"),_}return f.info(`Stored long-term memory: ${K.category} for entity ${K.entityId}`),K}async getLongTermMemories(J,Y,Q=10){let Z=this.getDb(),K=[k(H.agentId,this.runtime.agentId),k(H.entityId,J)];if(Y)K.push(k(H.category,Y));return(await Z.select().from(H).where(i(...K)).orderBy(JJ(H.confidence),JJ(H.updatedAt)).limit(Q)).map(($)=>({id:$.id,agentId:$.agentId,entityId:$.entityId,category:$.category,content:$.content,metadata:$.metadata,embedding:$.embedding,confidence:$.confidence,source:$.source,createdAt:$.createdAt,updatedAt:$.updatedAt,lastAccessedAt:$.lastAccessedAt,accessCount:$.accessCount}))}async updateLongTermMemory(J,Y,Q){let Z=this.getDb(),K={updatedAt:new Date};if(Q.content!==void 0)K.content=Q.content;if(Q.metadata!==void 0)K.metadata=Q.metadata;if(Q.confidence!==void 0)K.confidence=Q.confidence;if(Q.embedding!==void 0)K.embedding=Q.embedding;if(Q.lastAccessedAt!==void 0)K.lastAccessedAt=Q.lastAccessedAt;if(Q.accessCount!==void 0)K.accessCount=Q.accessCount;await Z.update(H).set(K).where(i(k(H.id,J),k(H.agentId,this.runtime.agentId),k(H.entityId,Y))),f.info(`Updated long-term memory: ${J} for entity ${Y}`)}async deleteLongTermMemory(J,Y){await this.getDb().delete(H).where(i(k(H.id,J),k(H.agentId,this.runtime.agentId),k(H.entityId,Y))),f.info(`Deleted long-term memory: ${J} for entity ${Y}`)}async getCurrentSessionSummary(J){let Q=await this.getDb().select().from(D).where(i(k(D.agentId,this.runtime.agentId),k(D.roomId,J))).orderBy(JJ(D.updatedAt)).limit(1);if(Q.length===0)return null;let Z=Q[0];return{id:Z.id,agentId:Z.agentId,roomId:Z.roomId,entityId:Z.entityId,summary:Z.summary,messageCount:Z.messageCount,lastMessageOffset:Z.lastMessageOffset,startTime:Z.startTime,endTime:Z.endTime,topics:Z.topics||[],metadata:Z.metadata,embedding:Z.embedding,createdAt:Z.createdAt,updatedAt:Z.updatedAt}}async storeSessionSummary(J){let Y=this.getDb(),Q=crypto.randomUUID(),Z=new Date,K={id:Q,createdAt:Z,updatedAt:Z,...J};return await Y.insert(D).values({id:K.id,agentId:K.agentId,roomId:K.roomId,entityId:K.entityId||null,summary:K.summary,messageCount:K.messageCount,lastMessageOffset:K.lastMessageOffset,startTime:K.startTime,endTime:K.endTime,topics:K.topics||[],metadata:K.metadata||{},embedding:K.embedding,createdAt:Z,updatedAt:Z}),f.info(`Stored session summary for room ${K.roomId}`),K}async updateSessionSummary(J,Y,Q){let Z=this.getDb(),K={updatedAt:new Date};if(Q.summary!==void 0)K.summary=Q.summary;if(Q.messageCount!==void 0)K.messageCount=Q.messageCount;if(Q.lastMessageOffset!==void 0)K.lastMessageOffset=Q.lastMessageOffset;if(Q.endTime!==void 0)K.endTime=Q.endTime;if(Q.topics!==void 0)K.topics=Q.topics;if(Q.metadata!==void 0)K.metadata=Q.metadata;if(Q.embedding!==void 0)K.embedding=Q.embedding;await Z.update(D).set(K).where(i(k(D.id,J),k(D.agentId,this.runtime.agentId),k(D.roomId,Y))),f.info(`Updated session summary: ${J} for room ${Y}`)}async getSessionSummaries(J,Y=5){return(await this.getDb().select().from(D).where(i(k(D.agentId,this.runtime.agentId),k(D.roomId,J))).orderBy(JJ(D.updatedAt)).limit(Y)).map((K)=>({id:K.id,agentId:K.agentId,roomId:K.roomId,entityId:K.entityId,summary:K.summary,messageCount:K.messageCount,lastMessageOffset:K.lastMessageOffset,startTime:K.startTime,endTime:K.endTime,topics:K.topics||[],metadata:K.metadata,embedding:K.embedding,createdAt:K.createdAt,updatedAt:K.updatedAt}))}async searchLongTermMemories(J,Y,Q=5,Z=0.7){if(!this.memoryConfig.longTermVectorSearchEnabled)return f.warn("Vector search is not enabled, falling back to recent memories"),this.getLongTermMemories(J,void 0,Q);let K=this.getDb();try{let _=Y.map((W)=>Number.isFinite(W)?Number(W.toFixed(6)):0),$=CJ`1 - (${tJ(H.embedding,_)})`,U=[k(H.agentId,this.runtime.agentId),k(H.entityId,J),CJ`${H.embedding} IS NOT NULL`];if(Z>0)U.push(eJ($,Z));return(await K.select({memory:H,similarity:$}).from(H).where(i(...U)).orderBy(JJ($)).limit(Q)).map((W)=>({id:W.memory.id,agentId:W.memory.agentId,entityId:W.memory.entityId,category:W.memory.category,content:W.memory.content,metadata:W.memory.metadata,embedding:W.memory.embedding,confidence:W.memory.confidence,source:W.memory.source,createdAt:W.memory.createdAt,updatedAt:W.memory.updatedAt,lastAccessedAt:W.memory.lastAccessedAt,accessCount:W.memory.accessCount,similarity:W.similarity}))}catch(_){return f.warn({error:_},"Vector search failed, falling back to recent memories"),this.getLongTermMemories(J,void 0,Q)}}async getFormattedLongTermMemories(J){let Y=await this.getLongTermMemories(J,void 0,20);if(Y.length===0)return"";let Q=new Map;for(let K of Y){if(!Q.has(K.category))Q.set(K.category,[]);Q.get(K.category)?.push(K)}let Z=[];for(let[K,_]of Q.entries()){let $=K.split("_").map((P)=>P.charAt(0).toUpperCase()+P.slice(1)).join(" "),U=_.map((P)=>`- ${P.content}`).join(`
2
+ `);Z.push(`**${$}**:
3
+ ${U}`)}return Z.join(`
4
4
 
5
- `)}}import{logger as R,ModelType as cJ,composePromptFromState as BJ}from"@elizaos/core";var dJ=`# Task: Summarize Conversation
5
+ `)}}import{logger as x,ModelType as JK,composePromptFromState as vJ}from"@elizaos/core";var KK=`# Task: Summarize Conversation
6
6
 
7
7
  You are analyzing a conversation to create a concise summary that captures the key points, topics, and important details.
8
8
 
@@ -31,7 +31,7 @@ Respond in this XML format:
31
31
  <point>First key point</point>
32
32
  <point>Second key point</point>
33
33
  </keyPoints>
34
- </summary>`,uJ=`# Task: Update and Condense Conversation Summary
34
+ </summary>`,QK=`# Task: Update and Condense Conversation Summary
35
35
 
36
36
  You are updating an existing conversation summary with new messages, while keeping the total summary concise.
37
37
 
@@ -62,10 +62,10 @@ Respond in this XML format:
62
62
  <point>First key point</point>
63
63
  <point>Second key point</point>
64
64
  </keyPoints>
65
- </summary>`;function gJ(J){let Q=J.match(/<text>([\s\S]*?)<\/text>/),Y=J.match(/<topics>([\s\S]*?)<\/topics>/),W=J.matchAll(/<point>([\s\S]*?)<\/point>/g),K=Q?Q[1].trim():"Summary not available",Z=Y?Y[1].split(",").map((U)=>U.trim()).filter(Boolean):[],$=Array.from(W).map((U)=>U[1].trim());return{summary:K,topics:Z,keyPoints:$}}var OJ={name:"MEMORY_SUMMARIZATION",description:"Automatically summarizes conversations to optimize context usage",similes:["CONVERSATION_SUMMARY","CONTEXT_COMPRESSION","MEMORY_OPTIMIZATION"],alwaysRun:!0,validate:async(J,Q)=>{if(!Q.content?.text)return R.debug("Skipping summarization: no message text"),!1;let Y=J.getService("memory");if(!Y)return R.debug("Skipping summarization: memory service not available"),!1;let W=Y.getConfig(),K=await J.countMemories(Q.roomId,!1,"messages"),Z=K>=W.shortTermSummarizationThreshold;return R.debug({roomId:Q.roomId,currentMessageCount:K,threshold:W.shortTermSummarizationThreshold,shouldSummarize:Z},"Summarization validation check"),Z},handler:async(J,Q)=>{let Y=J.getService("memory");if(!Y){R.error("MemoryService not found");return}let W=Y.getConfig(),{roomId:K}=Q;try{R.info(`Starting summarization for room ${K}`);let Z=await Y.getCurrentSessionSummary(K),$=Z?.lastMessageOffset||0,U=await J.countMemories(K,!1,"messages"),V=await J.getMemories({tableName:"messages",roomId:K,count:W.shortTermSummarizationThreshold,unique:!1,start:$});if(V.length===0){R.debug("No new messages to summarize");return}let _=V.sort((C,z)=>(C.createdAt||0)-(z.createdAt||0)),B=_.map((C)=>{return`${C.entityId===J.agentId?J.character.name:"User"}: ${C.content.text||"[non-text message]"}`}).join(`
66
- `),O=await J.composeState(Q),k,G;if(Z)G=uJ,k=BJ({state:{...O,existingSummary:Z.summary,existingTopics:Z.topics?.join(", ")||"None",newMessages:B},template:G});else G=dJ,k=BJ({state:{...O,recentMessages:B},template:G});let j=await J.useModel(cJ.TEXT_LARGE,{prompt:k,maxTokens:W.summaryMaxTokens||2500}),A=gJ(j);R.info(`${Z?"Updated":"Generated"} summary: ${A.summary.substring(0,100)}...`);let N=U,F=_[0],E=_[_.length-1],b=Z?Z.startTime:F?.createdAt&&F.createdAt>0?new Date(F.createdAt):new Date,h=E?.createdAt&&E.createdAt>0?new Date(E.createdAt):new Date;if(Z)await Y.updateSessionSummary(Z.id,{summary:A.summary,messageCount:Z.messageCount+_.length,lastMessageOffset:N,endTime:h,topics:A.topics,metadata:{keyPoints:A.keyPoints}}),R.info(`Updated summary for room ${K}: ${_.length} new messages processed (offset: ${$} → ${N})`);else await Y.storeSessionSummary({agentId:J.agentId,roomId:K,entityId:Q.entityId!==J.agentId?Q.entityId:void 0,summary:A.summary,messageCount:_.length,lastMessageOffset:N,startTime:b,endTime:h,topics:A.topics,metadata:{keyPoints:A.keyPoints}}),R.info(`Created new summary for room ${K}: ${_.length} messages summarized (offset: 0 → ${N})`)}catch(Z){R.error({error:Z},"Error during summarization:")}},examples:[]};import{logger as P,ModelType as iJ,composePromptFromState as nJ}from"@elizaos/core";var KJ;((_)=>{_.IDENTITY="identity";_.EXPERTISE="expertise";_.PROJECTS="projects";_.PREFERENCES="preferences";_.DATA_SOURCES="data_sources";_.GOALS="goals";_.CONSTRAINTS="constraints";_.DEFINITIONS="definitions";_.BEHAVIORAL_PATTERNS="behavioral_patterns"})(KJ||={});var aJ=`# Task: Extract Long-Term Memory
65
+ </summary>`;function YK(J){let Y=J.match(/<text>([\s\S]*?)<\/text>/),Q=J.match(/<topics>([\s\S]*?)<\/topics>/),Z=J.matchAll(/<point>([\s\S]*?)<\/point>/g),K=Y?Y[1].trim():"Summary not available",_=Q?Q[1].split(",").map((U)=>U.trim()).filter(Boolean):[],$=Array.from(Z).map((U)=>U[1].trim());return{summary:K,topics:_,keyPoints:$}}var TJ={name:"MEMORY_SUMMARIZATION",description:"Automatically summarizes conversations to optimize context usage",similes:["CONVERSATION_SUMMARY","CONTEXT_COMPRESSION","MEMORY_OPTIMIZATION"],alwaysRun:!0,validate:async(J,Y)=>{if(!Y.content?.text)return x.debug("Skipping summarization: no message text"),!1;let Q=J.getService("memory");if(!Q)return x.debug("Skipping summarization: memory service not available"),!1;let Z=Q.getConfig(),K=await J.countMemories(Y.roomId,!1,"messages"),_=await Q.getCurrentSessionSummary(Y.roomId);if(!_){let $=K>=Z.shortTermSummarizationThreshold;return x.debug({roomId:Y.roomId,currentMessageCount:K,threshold:Z.shortTermSummarizationThreshold,shouldSummarize:$,reason:"initial_summary_check"},"Summarization validation check"),$}else{let $=K-_.lastMessageOffset,U=$>=Z.shortTermSummarizationInterval;return x.debug({roomId:Y.roomId,currentMessageCount:K,lastOffset:_.lastMessageOffset,newMessageCount:$,interval:Z.shortTermSummarizationInterval,shouldUpdate:U,reason:"summary_update_check"},"Summarization validation check"),U}},handler:async(J,Y)=>{let Q=J.getService("memory");if(!Q){x.error("MemoryService not found");return}let Z=Q.getConfig(),{roomId:K}=Y;try{x.info(`Starting summarization for room ${K}`);let _=await Q.getCurrentSessionSummary(K),$=_?.lastMessageOffset||0,U=await J.countMemories(K,!1,"messages"),P=U-$,W=Z.summaryMaxNewMessages||50,R=Math.min(P,W);if(R===0){x.debug("No new messages to summarize");return}if(P>W)x.warn(`Capping new messages at ${W} (${P} available). Oldest messages will be skipped.`);let T=await J.getMemories({tableName:"messages",roomId:K,count:R,unique:!1,start:$});if(T.length===0){x.debug("No new messages retrieved");return}let O=T.sort((c,h)=>(c.createdAt||0)-(h.createdAt||0)),S=O.map((c)=>{return`${c.entityId===J.agentId?J.character.name:"User"}: ${c.content.text||"[non-text message]"}`}).join(`
66
+ `),E=await J.composeState(Y),B,b;if(_)b=QK,B=vJ({state:{...E,existingSummary:_.summary,existingTopics:_.topics?.join(", ")||"None",newMessages:S},template:b});else b=KK,B=vJ({state:{...E,recentMessages:S},template:b});let l=await J.useModel(JK.TEXT_LARGE,{prompt:B,maxTokens:Z.summaryMaxTokens||2500}),N=YK(l);x.info(`${_?"Updated":"Generated"} summary: ${N.summary.substring(0,100)}...`);let I=U,C=O[0],p=O[O.length-1],M=_?_.startTime:C?.createdAt&&C.createdAt>0?new Date(C.createdAt):new Date,w=p?.createdAt&&p.createdAt>0?new Date(p.createdAt):new Date;if(_)await Q.updateSessionSummary(_.id,K,{summary:N.summary,messageCount:_.messageCount+O.length,lastMessageOffset:I,endTime:w,topics:N.topics,metadata:{keyPoints:N.keyPoints}}),x.info(`Updated summary for room ${K}: ${O.length} new messages processed (offset: ${$} → ${I})`);else await Q.storeSessionSummary({agentId:J.agentId,roomId:K,entityId:Y.entityId!==J.agentId?Y.entityId:void 0,summary:N.summary,messageCount:O.length,lastMessageOffset:I,startTime:M,endTime:w,topics:N.topics,metadata:{keyPoints:N.keyPoints}}),x.info(`Created new summary for room ${K}: ${O.length} messages summarized (offset: 0 → ${I})`)}catch(_){x.error({error:_},"Error during summarization:")}},examples:[]};import{logger as v,ModelType as ZK,composePromptFromState as _K}from"@elizaos/core";var EJ;((Z)=>{Z.EPISODIC="episodic";Z.SEMANTIC="semantic";Z.PROCEDURAL="procedural"})(EJ||={});var $K=`# Task: Extract Long-Term Memory (Strict Criteria)
67
67
 
68
- You are analyzing a conversation to extract facts that should be remembered long-term about the user.
68
+ You are analyzing a conversation to extract ONLY the most critical, persistent information about the user using cognitive science memory categories.
69
69
 
70
70
  # Recent Messages
71
71
  {{recentMessages}}
@@ -73,50 +73,150 @@ You are analyzing a conversation to extract facts that should be remembered long
73
73
  # Current Long-Term Memories
74
74
  {{existingMemories}}
75
75
 
76
- # Memory Categories
77
- 1. **identity**: User's name, role, identity (e.g., "I'm a data scientist")
78
- 2. **expertise**: User's skills, knowledge domains, or unfamiliarity with topics
79
- 3. **projects**: Ongoing projects, past interactions, recurring topics
80
- 4. **preferences**: Communication style, format preferences, verbosity, etc.
81
- 5. **data_sources**: Frequently used files, databases, APIs
82
- 6. **goals**: Broader intentions (e.g., "preparing for interview")
83
- 7. **constraints**: User-defined rules or limitations
84
- 8. **definitions**: Custom terms, acronyms, glossaries
85
- 9. **behavioral_patterns**: How the user tends to interact
76
+ # Memory Categories (Based on Cognitive Science)
77
+
78
+ ## 1. EPISODIC Memory
79
+ Personal experiences and specific events with temporal/spatial context.
80
+ **Examples:**
81
+ - "User completed migration project from MongoDB to PostgreSQL in Q2 2024"
82
+ - "User encountered authentication bug in production on March 15th"
83
+ - "User had a negative experience with Docker networking in previous job"
84
+
85
+ **Requirements:**
86
+ - Must include WHO did WHAT, WHEN/WHERE
87
+ - Must be a specific, concrete event (not a pattern)
88
+ - Must have significant impact or relevance to future work
89
+
90
+ ## 2. SEMANTIC Memory
91
+ General facts, concepts, knowledge, and established truths about the user.
92
+ **Examples:**
93
+ - "User is a senior backend engineer with 8 years experience"
94
+ - "User specializes in distributed systems and microservices architecture"
95
+ - "User's primary programming language is TypeScript"
96
+ - "User works at Acme Corp as technical lead"
97
+
98
+ **Requirements:**
99
+ - Must be factual, timeless information
100
+ - Must be explicitly stated or demonstrated conclusively
101
+ - No speculation or inference from single instances
102
+ - Core identity, expertise, or knowledge only
103
+
104
+ ## 3. PROCEDURAL Memory
105
+ Skills, workflows, methodologies, and how-to knowledge.
106
+ **Examples:**
107
+ - "User follows strict TDD workflow: write tests first, then implementation"
108
+ - "User prefers git rebase over merge to maintain linear history"
109
+ - "User's debugging process: check logs → reproduce locally → binary search"
110
+ - "User always writes JSDoc comments before implementing functions"
111
+
112
+ **Requirements:**
113
+ - Must describe HOW user does something
114
+ - Must be a repeated, consistent pattern (seen 3+ times or explicitly stated as standard practice)
115
+ - Must be a workflow, methodology, or skill application
116
+ - Not one-off preferences
117
+
118
+ # ULTRA-STRICT EXTRACTION CRITERIA
119
+
120
+ ## ✅ DO EXTRACT (Only These):
121
+
122
+ **EPISODIC:**
123
+ - Significant completed projects or milestones
124
+ - Important bugs, incidents, or problems encountered
125
+ - Major decisions made with lasting impact
126
+ - Formative experiences that shape future work
127
+
128
+ **SEMANTIC:**
129
+ - Professional identity (role, title, company)
130
+ - Core expertise and specializations (stated explicitly or demonstrated conclusively)
131
+ - Primary languages, frameworks, or tools (not exploratory use)
132
+ - Established facts about their work context
133
+
134
+ **PROCEDURAL:**
135
+ - Consistent workflows demonstrated 3+ times or explicitly stated
136
+ - Standard practices user always follows
137
+ - Methodology preferences with clear rationale
138
+ - Debugging, testing, or development processes
139
+
140
+ ## ❌ NEVER EXTRACT:
141
+
142
+ - **One-time requests or tasks** (e.g., "can you generate an image", "help me debug this")
143
+ - **Casual conversations** without lasting significance
144
+ - **Exploratory questions** (e.g., "how does X work?")
145
+ - **Temporary context** (current bug, today's task)
146
+ - **Preferences from single occurrence** (e.g., user asked for code once)
147
+ - **Social pleasantries** (thank you, greetings)
148
+ - **Testing or experimentation** (trying out a feature)
149
+ - **Common patterns everyone has** (likes clear explanations)
150
+ - **Situational information** (working on feature X today)
151
+ - **Opinions without persistence** (single complaint, isolated praise)
152
+ - **General knowledge** (not specific to user)
153
+
154
+ # Quality Gates (ALL Must Pass)
155
+
156
+ 1. **Significance Test**: Will this matter in 3+ months?
157
+ 2. **Specificity Test**: Is this concrete and actionable?
158
+ 3. **Evidence Test**: Is there strong evidence (3+ instances OR explicit self-identification)?
159
+ 4. **Uniqueness Test**: Is this specific to THIS user (not generic)?
160
+ 5. **Confidence Test**: Confidence must be >= 0.85 (be VERY conservative)
161
+ 6. **Non-Redundancy Test**: Does this add NEW information not in existing memories?
162
+
163
+ # Confidence Scoring (Be Conservative)
164
+
165
+ - **0.95-1.0**: User explicitly stated as core identity/practice AND demonstrated multiple times
166
+ - **0.85-0.94**: User explicitly stated OR consistently demonstrated 5+ times
167
+ - **0.75-0.84**: Strong pattern (3-4 instances) with supporting context
168
+ - **Below 0.75**: DO NOT EXTRACT (insufficient evidence)
169
+
170
+ # Critical Instructions
171
+
172
+ 1. **Default to NOT extracting** - When in doubt, skip it
173
+ 2. **Require overwhelming evidence** - One or two mentions is NOT enough
174
+ 3. **Focus on what's PERSISTENT** - Not what's temporary or situational
175
+ 4. **Verify against existing memories** - Don't duplicate or contradict
176
+ 5. **Maximum 2-3 extractions per run** - Quality over quantity
177
+
178
+ **If there are no qualifying facts (which is common), respond with <memories></memories>**
179
+
180
+ # Response Format
86
181
 
87
- # Instructions
88
- Extract any NEW information that should be remembered long-term. For each item:
89
- - Determine which category it belongs to
90
- - Write a clear, factual statement
91
- - Assess confidence (0.0 to 1.0)
92
- - Only include information explicitly stated or strongly implied
93
-
94
- If there are no new long-term facts to extract, respond with <memories></memories>
95
-
96
- Respond in this XML format:
97
182
  <memories>
98
183
  <memory>
99
- <category>identity</category>
100
- <content>User is a software engineer specializing in backend development</content>
184
+ <category>semantic</category>
185
+ <content>User is a senior TypeScript developer with 8 years of backend experience</content>
101
186
  <confidence>0.95</confidence>
102
187
  </memory>
103
188
  <memory>
104
- <category>preferences</category>
105
- <content>Prefers code examples over lengthy explanations</content>
106
- <confidence>0.85</confidence>
189
+ <category>procedural</category>
190
+ <content>User follows TDD workflow: writes tests before implementation, runs tests after each change</content>
191
+ <confidence>0.88</confidence>
107
192
  </memory>
108
- </memories>`;function oJ(J){let Q=J.matchAll(/<memory>[\s\S]*?<category>(.*?)<\/category>[\s\S]*?<content>(.*?)<\/content>[\s\S]*?<confidence>(.*?)<\/confidence>[\s\S]*?<\/memory>/g),Y=[];for(let W of Q){let K=W[1].trim(),Z=W[2].trim(),$=parseFloat(W[3].trim());if(!Object.values(KJ).includes(K)){P.warn(`Invalid memory category: ${K}`);continue}if(Z&&!isNaN($))Y.push({category:K,content:Z,confidence:$})}return Y}var NJ={name:"LONG_TERM_MEMORY_EXTRACTION",description:"Extracts long-term facts about users from conversations",similes:["MEMORY_EXTRACTION","FACT_LEARNING","USER_PROFILING"],alwaysRun:!0,validate:async(J,Q)=>{if(P.debug(`Validating long-term memory extraction for message: ${Q.content?.text}`),Q.entityId===J.agentId)return P.debug("Skipping long-term memory extraction for agent's own message"),!1;if(!Q.content?.text)return P.debug("Skipping long-term memory extraction for message without text"),!1;let Y=J.getService("memory");if(!Y)return P.debug("MemoryService not found"),!1;if(!Y.getConfig().longTermExtractionEnabled)return P.debug("Long-term memory extraction is disabled"),!1;let K=await J.countMemories(Q.roomId,!1,"messages"),Z=await Y.shouldRunExtraction(Q.entityId,Q.roomId,K);return P.debug(`Should run extraction: ${Z}`),Z},handler:async(J,Q)=>{let Y=J.getService("memory");if(!Y){P.error("MemoryService not found");return}let W=Y.getConfig(),{entityId:K,roomId:Z}=Q;try{P.info(`Extracting long-term memories for entity ${K}`);let U=(await J.getMemories({tableName:"messages",roomId:Z,count:20,unique:!1})).sort((A,N)=>(A.createdAt||0)-(N.createdAt||0)).map((A)=>{return`${A.entityId===J.agentId?J.character.name:"User"}: ${A.content.text||"[non-text message]"}`}).join(`
109
- `),V=await Y.getLongTermMemories(K,void 0,30),_=V.length>0?V.map((A)=>`[${A.category}] ${A.content} (confidence: ${A.confidence})`).join(`
110
- `):"None yet",B=await J.composeState(Q),O=nJ({state:{...B,recentMessages:U,existingMemories:_},template:aJ}),k=await J.useModel(iJ.TEXT_LARGE,{prompt:O}),G=oJ(k);P.info(`Extracted ${G.length} long-term memories`);for(let A of G)if(A.confidence>=W.longTermConfidenceThreshold)await Y.storeLongTermMemory({agentId:J.agentId,entityId:K,category:A.category,content:A.content,confidence:A.confidence,source:"conversation",metadata:{roomId:Z,extractedAt:new Date().toISOString()}}),P.info(`Stored long-term memory: [${A.category}] ${A.content.substring(0,50)}...`);else P.debug(`Skipped low-confidence memory: ${A.content} (confidence: ${A.confidence})`);let j=await J.countMemories(Z,!1,"messages");await Y.setLastExtractionCheckpoint(K,Z,j),P.debug(`Updated extraction checkpoint to ${j} for entity ${K} in room ${Z}`)}catch($){P.error({error:$},"Error during long-term memory extraction:")}},examples:[]};import{addHeader as f,ChannelType as i,formatMessages as QJ,formatPosts as FJ,getEntityDetails as EJ,logger as mJ}from"@elizaos/core";var kJ={name:"SHORT_TERM_MEMORY",description:"Adaptive conversation context with smart summarization",position:95,get:async(J,Q,Y)=>{try{let W=J.getService("memory");if(!W)return{data:{summaries:[],recentMessages:[],mode:"disabled"},values:{},text:""};let{roomId:K}=Q,Z=W.getConfig();if(await J.countMemories(K,!1,"messages")<Z.shortTermSummarizationThreshold){let U=J.getConversationLength(),[V,_,B]=await Promise.all([EJ({runtime:J,roomId:K}),J.getRoom(K),J.getMemories({tableName:"messages",roomId:K,count:U,unique:!1})]),O=B.filter((L)=>L.content?.type==="action_result"&&L.metadata?.type==="action_result"),k=B.filter((L)=>!(L.content?.type==="action_result"&&L.metadata?.type==="action_result")),G=_?.type?_.type===i.FEED||_.type===i.THREAD:!1,[j,A]=await Promise.all([QJ({messages:k,entities:V}),FJ({messages:k,entities:V,conversationHeader:!1})]),N="";if(O.length>0){let L=new Map;for(let S of O){let v=String(S.content?.runId||"unknown");if(!L.has(v))L.set(v,[]);L.get(v)?.push(S)}let x=Array.from(L.entries()).slice(-3).map(([S,v])=>{let WJ=v.sort((I,M)=>(I.createdAt||0)-(M.createdAt||0)),YJ=WJ[0]?.content?.planThought||"",HJ=WJ.map((I)=>{let M=I.content?.actionName||"Unknown",PJ=I.content?.actionStatus||"unknown",ZJ=I.content?.planStep||"",o=I.content?.text||"",_J=I.content?.error||"",y=` - ${M} (${PJ})`;if(ZJ)y+=` [${ZJ}]`;if(_J)y+=`: Error - ${_J}`;else if(o&&o!==`Executed action: ${M}`)y+=`: ${o}`;return y}).join(`
111
- `);return`**Action Run ${S.slice(0,8)}**${YJ?` - "${YJ}"`:""}
112
- ${HJ}`}).join(`
113
-
114
- `);N=x?f("# Recent Action Executions",x):""}let F=A&&A.length>0?f("# Posts in Thread",A):"",E=j&&j.length>0?f("# Conversation Messages",j):"";if(!F&&!E&&k.length===0&&!Q.content.text)return{data:{summaries:[],recentMessages:[],actionResults:[],mode:"full_conversation"},values:{recentMessage:"No recent message available."},text:"No recent messages available"};let b="No recent message available.";if(k.length>0){let L=[...k].sort((S,v)=>(v.createdAt||0)-(S.createdAt||0))[0],x=QJ({messages:[L],entities:V});if(x)b=x}let h=Q.metadata,C=V.find((L)=>L.id===Q.entityId)?.names[0]||h?.entityName||"Unknown User",z=Q.content.text,T=!!z?.trim(),n=T?f("# Received Message",`${C}: ${z}`):"",a=T?f("# Focus your response",`You are replying to the above message from **${C}**. Keep your answer relevant to that message. Do not repeat earlier replies unless the sender asks again.`):"",qJ=[G?F:E,N,E||F||Q.content.text?n:"",E||F||Q.content.text?a:""].filter(Boolean).join(`
115
-
116
- `);return{data:{summaries:[],recentMessages:k,actionResults:O,mode:"full_conversation"},values:{...(G?F:E)&&{recentMessages:G?F:E},...F&&{recentPosts:F},...N&&{recentActionResults:N},...b&&{recentMessage:b},...n&&{receivedMessageHeader:n},...a&&{focusHeader:a}},text:qJ}}else{let U=await W.getCurrentSessionSummary(K),V=U?.lastMessageOffset||0,_=await J.getMemories({tableName:"messages",roomId:K,count:Z.shortTermRetainRecent,unique:!1,start:V}),B=await EJ({runtime:J,roomId:K}),O=await J.getRoom(K),k=O?.type?O.type===i.FEED||O.type===i.THREAD:!1,G="";if(_.length>0){let z=_.filter((T)=>!(T.content?.type==="action_result"&&T.metadata?.type==="action_result"));if(k)G=FJ({messages:z,entities:B,conversationHeader:!1});else G=QJ({messages:z,entities:B});if(G)G=f("# Recent Messages",G)}let j="";if(U){let z=`${U.messageCount} messages`,T=new Date(U.startTime).toLocaleDateString();if(j=`**Previous Conversation** (${z}, ${T})
117
- `,j+=U.summary,U.topics&&U.topics.length>0)j+=`
118
- *Topics: ${U.topics.join(", ")}*`;j=f("# Conversation Summary",j)}let A=Q.metadata,N=B.find((z)=>z.id===Q.entityId)?.names[0]||A?.entityName||"Unknown User",F=Q.content.text,E=!!F?.trim(),b=E?f("# Received Message",`${N}: ${F}`):"",h=E?f("# Focus your response",`You are replying to the above message from **${N}**. Keep your answer relevant to that message.`):"",C=[j,G,E?b:"",E?h:""].filter(Boolean).join(`
119
-
120
- `);return{data:{summaries:U?[U]:[],recentMessages:_,mode:"summarized"},values:{...j&&{sessionSummaries:j},...G&&{recentMessages:G},...b&&{receivedMessageHeader:b},...h&&{focusHeader:h}},text:C}}}catch(W){return mJ.error({error:W},"Error in shortTermMemoryProvider:"),{data:{summaries:[],recentMessages:[],mode:"error"},values:{},text:"Error retrieving conversation context."}}}};import{logger as sJ,addHeader as tJ}from"@elizaos/core";var LJ={name:"LONG_TERM_MEMORY",description:"Persistent facts and preferences about the user",position:50,get:async(J,Q,Y)=>{try{let W=J.getService("memory");if(!W)return{data:{memories:[]},values:{longTermMemories:""},text:""};let{entityId:K}=Q;if(K===J.agentId)return{data:{memories:[]},values:{longTermMemories:""},text:""};let Z=await W.getLongTermMemories(K,void 0,25);if(Z.length===0)return{data:{memories:[]},values:{longTermMemories:""},text:""};let $=await W.getFormattedLongTermMemories(K),U=tJ("# What I Know About You",$),V=new Map;for(let B of Z){let O=V.get(B.category)||0;V.set(B.category,O+1)}let _=Array.from(V.entries()).map(([B,O])=>`${B}: ${O}`).join(", ");return{data:{memories:Z,categoryCounts:Object.fromEntries(V)},values:{longTermMemories:U,memoryCategories:_},text:U}}catch(W){return sJ.error({error:W},"Error in longTermMemoryProvider:"),{data:{memories:[]},values:{longTermMemories:""},text:""}}}};var rJ={name:"memory",description:"Advanced memory management with conversation summarization and long-term persistent memory",services:[w],evaluators:[OJ,NJ],providers:[LJ,kJ],schema:JJ},eJ=rJ;export{q as sessionSummaries,rJ as memoryPlugin,XJ as memoryAccessLogs,X as longTermMemories,eJ as default,w as MemoryService,KJ as LongTermMemoryCategory};
121
-
122
- //# debugId=480869ED6980B15B64756E2164756E21
193
+ <memory>
194
+ <category>episodic</category>
195
+ <content>User led database migration from MongoDB to PostgreSQL for payment system in Q2 2024</content>
196
+ <confidence>0.92</confidence>
197
+ </memory>
198
+ </memories>`;function VK(J){let Y=J.matchAll(/<memory>[\s\S]*?<category>(.*?)<\/category>[\s\S]*?<content>(.*?)<\/content>[\s\S]*?<confidence>(.*?)<\/confidence>[\s\S]*?<\/memory>/g),Q=[];for(let Z of Y){let K=Z[1].trim(),_=Z[2].trim(),$=parseFloat(Z[3].trim());if(!Object.values(EJ).includes(K)){v.warn(`Invalid memory category: ${K}`);continue}if(_&&!isNaN($))Q.push({category:K,content:_,confidence:$})}return Q}var SJ={name:"LONG_TERM_MEMORY_EXTRACTION",description:"Extracts long-term facts about users from conversations",similes:["MEMORY_EXTRACTION","FACT_LEARNING","USER_PROFILING"],alwaysRun:!0,validate:async(J,Y)=>{if(v.debug(`Validating long-term memory extraction for message: ${Y.content?.text}`),Y.entityId===J.agentId)return v.debug("Skipping long-term memory extraction for agent's own message"),!1;if(!Y.content?.text)return v.debug("Skipping long-term memory extraction for message without text"),!1;let Q=J.getService("memory");if(!Q)return v.debug("MemoryService not found"),!1;if(!Q.getConfig().longTermExtractionEnabled)return v.debug("Long-term memory extraction is disabled"),!1;let K=await J.countMemories(Y.roomId,!1,"messages"),_=await Q.shouldRunExtraction(Y.entityId,Y.roomId,K);return v.debug(`Should run extraction: ${_}`),_},handler:async(J,Y)=>{let Q=J.getService("memory");if(!Q){v.error("MemoryService not found");return}let Z=Q.getConfig(),{entityId:K,roomId:_}=Y;try{v.info(`Extracting long-term memories for entity ${K}`);let U=(await J.getMemories({tableName:"messages",roomId:_,count:20,unique:!1})).sort((B,b)=>(B.createdAt||0)-(b.createdAt||0)).map((B)=>{return`${B.entityId===J.agentId?J.character.name:"User"}: ${B.content.text||"[non-text message]"}`}).join(`
199
+ `),P=await Q.getLongTermMemories(K,void 0,30),W=P.length>0?P.map((B)=>`[${B.category}] ${B.content} (confidence: ${B.confidence})`).join(`
200
+ `):"None yet",R=await J.composeState(Y),T=_K({state:{...R,recentMessages:U,existingMemories:W},template:$K}),O=await J.useModel(ZK.TEXT_LARGE,{prompt:T}),S=VK(O);v.info(`Extracted ${S.length} long-term memories`);for(let B of S)if(B.confidence>=Math.max(Z.longTermConfidenceThreshold,0.85))await Q.storeLongTermMemory({agentId:J.agentId,entityId:K,category:B.category,content:B.content,confidence:B.confidence,source:"conversation",metadata:{roomId:_,extractedAt:new Date().toISOString()}}),v.info(`Stored long-term memory: [${B.category}] ${B.content.substring(0,50)}...`);else v.debug(`Skipped low-confidence memory: ${B.content} (confidence: ${B.confidence}, threshold: ${Math.max(Z.longTermConfidenceThreshold,0.85)})`);let E=await J.countMemories(_,!1,"messages");await Q.setLastExtractionCheckpoint(K,_,E),v.debug(`Updated extraction checkpoint to ${E} for entity ${K} in room ${_}`)}catch($){v.error({error:$},"Error during long-term memory extraction:")}},examples:[]};import{addHeader as y,ChannelType as XJ,formatMessages as jJ,formatPosts as BJ,getEntityDetails as hJ,logger as WK}from"@elizaos/core";var xJ=async(J,Y,Q,Z)=>{let K=await J.getRoomsForParticipants([Y,Q]);return J.getMemoriesByRoomIds({tableName:"messages",roomIds:K.filter((_)=>_!==Z),limit:20})},IJ={name:"SHORT_TERM_MEMORY",description:"Unified conversation context with smart summarization and interactions",position:95,get:async(J,Y,Q)=>{try{let Z=J.getService("memory"),{roomId:K}=Y,_=16,$=Z?.getConfig()||{shortTermSummarizationThreshold:16,shortTermRetainRecent:6},U=await J.countMemories(K,!1,"messages");if(!(Z&&U>=$.shortTermSummarizationThreshold)){let[W,R,T,O]=await Promise.all([hJ({runtime:J,roomId:K}),J.getRoom(K),J.getMemories({tableName:"messages",roomId:K,count:16,unique:!1}),Y.entityId!==J.agentId?xJ(J,Y.entityId,J.agentId,K):Promise.resolve([])]),S=T.filter((X)=>X.content?.type==="action_result"&&X.metadata?.type==="action_result"),E=T.filter((X)=>!(X.content?.type==="action_result"&&X.metadata?.type==="action_result")),B=R?.type?R.type===XJ.FEED||R.type===XJ.THREAD:!1,[b,l]=await Promise.all([jJ({messages:E,entities:W}),BJ({messages:E,entities:W,conversationHeader:!1})]),N=(X,A)=>{return X.sort((V,G)=>(V.createdAt||0)-(G.createdAt||0)).map((V)=>{let j=W.find((u)=>u.id===V.entityId)?.names[0]||(V.entityId===J.agentId?J.character.name:"Unknown"),L=V.createdAt?new Date(V.createdAt).toLocaleString():"Unknown time",z=V.content.text||"",q=A&&V.content.internalMonologue?`
201
+ [Internal thought: ${V.content.internalMonologue}]`:"";return`[${L}] ${j}: ${z}${q}`}).join(`
202
+ `)},I=N(E,!1),C=N(E,!0),p="";if(S.length>0){let X=new Map;for(let V of S){let G=String(V.content?.runId||"unknown");if(!X.has(G))X.set(G,[]);let j=X.get(G);if(j)j.push(V)}let A=Array.from(X.entries()).slice(-3).map(([V,G])=>{let j=G.sort((q,u)=>(q.createdAt||0)-(u.createdAt||0)),L=j[0]?.content?.planThought||"",z=j.map((q)=>{let u=q.content?.actionName||"Unknown",_J=q.content?.actionStatus||"unknown",m=q.content?.planStep||"",o=q.content?.text||"",g=q.content?.error||"",$J=` - ${u} (${_J})`;if(m)$J+=` [${m}]`;if(g)$J+=`: Error - ${g}`;else if(o&&o!==`Executed action: ${u}`)$J+=`: ${o}`;return $J}).join(`
203
+ `);return`**Action Run ${V.slice(0,8)}**${L?` - "${L}"`:""}
204
+ ${z}`}).join(`
205
+
206
+ `);p=A?y("# Recent Action Executions",A):""}let M=l&&l.length>0?y("# Posts in Thread",l):"",w=b&&b.length>0?y("# Conversation Messages",b):"";if(!M&&!w&&E.length===0&&!Y.content.text)return{data:{summaries:[],recentMessages:[],recentInteractions:[],actionResults:[],mode:"full_conversation"},values:{recentPosts:"",recentMessages:"",recentMessageInteractions:"",recentPostInteractions:"",recentInteractions:"",recentActionResults:"",recentMessage:"No recent message available.",conversationLog:"",conversationLogWithAgentThoughts:""},text:"No recent messages available"};let c="No recent message available.";if(E.length>0){let X=[...E].sort((V,G)=>(G.createdAt||0)-(V.createdAt||0))[0],A=jJ({messages:[X],entities:W});if(A)c=A}let h=Y.metadata,QJ=W.find((X)=>X.id===Y.entityId)?.names[0]||h?.entityName||"Unknown User",s=Y.content.text,r=!!s?.trim(),n=r?y("# Received Message",`${QJ}: ${s}`):"",t=r?y("# Focus your response",`You are replying to the above message from **${QJ}**. Keep your answer relevant to that message. Do not repeat earlier replies unless the sender asks again.`):"",d=new Map;if(O.length>0){let X=[...new Set(O.map((j)=>j.entityId).filter((j)=>j!==J.agentId))],A=new Set(X),V=new Set;W.forEach((j)=>{if(A.has(j.id))d.set(j.id,j),V.add(j.id)});let G=X.filter((j)=>!V.has(j));if(G.length>0)(await Promise.all(G.map((L)=>J.getEntityById(L)))).forEach((L,z)=>{if(L)d.set(G[z],L)})}let a=async(X)=>{return X.map((V)=>{let G=V.entityId===J.agentId,j;if(G)j=J.character.name;else j=d.get(V.entityId)?.metadata?.userName||"unknown";return`${j}: ${V.content.text}`}).join(`
207
+ `)},UJ=async(X,A)=>{let V=[...A],G=new Set(A.map((L)=>L.id));for(let[L,z]of d.entries())if(!G.has(L))V.push(z);return BJ({messages:X,entities:V,conversationHeader:!0})},[YJ,e]=await Promise.all([a(O),UJ(O,W)]),ZJ={summaries:[],recentMessages:E,recentInteractions:O,actionResults:S,mode:"full_conversation"},FJ={recentPosts:M,recentMessages:w,recentMessageInteractions:YJ,recentPostInteractions:e,recentInteractions:B?e:YJ,recentActionResults:p,recentMessage:c,conversationLog:I,conversationLogWithAgentThoughts:C},F=[B?M:w,p,w||M||Y.content.text?n:"",w||M||Y.content.text?t:""].filter(Boolean).join(`
208
+
209
+ `);return{data:ZJ,values:FJ,text:F}}else{let W=await Z.getCurrentSessionSummary(K),R=W?.lastMessageOffset||0,T=$.shortTermRetainRecent,[O,S,E,B]=await Promise.all([hJ({runtime:J,roomId:K}),J.getRoom(K),J.getMemories({tableName:"messages",roomId:K,count:T,unique:!1,start:R}),Y.entityId!==J.agentId?xJ(J,Y.entityId,J.agentId,K):Promise.resolve([])]),b=S?.type?S.type===XJ.FEED||S.type===XJ.THREAD:!1,l=E.filter((F)=>F.content?.type==="action_result"&&F.metadata?.type==="action_result"),N=E.filter((F)=>!(F.content?.type==="action_result"&&F.metadata?.type==="action_result")),I="";if(l.length>0){let F=new Map;for(let A of l){let V=String(A.content?.runId||"unknown");if(!F.has(V))F.set(V,[]);F.get(V)?.push(A)}let X=Array.from(F.entries()).slice(-3).map(([A,V])=>{let G=V.sort((z,q)=>(z.createdAt||0)-(q.createdAt||0)),j=G[0]?.content?.planThought||"",L=G.map((z)=>{let q=z.content?.actionName||"Unknown",u=z.content?.actionStatus||"unknown",_J=z.content?.planStep||"",m=z.content?.text||"",o=z.content?.error||"",g=` - ${q} (${u})`;if(_J)g+=` [${_J}]`;if(o)g+=`: Error - ${o}`;else if(m&&m!==`Executed action: ${q}`)g+=`: ${m}`;return g}).join(`
210
+ `);return`**Action Run ${A.slice(0,8)}**${j?` - "${j}"`:""}
211
+ ${L}`}).join(`
212
+
213
+ `);I=X?y("# Recent Action Executions",X):""}let C="";if(N.length>0){if(b)C=BJ({messages:N,entities:O,conversationHeader:!1});else C=jJ({messages:N,entities:O});if(C)C=y("# Recent Messages",C)}let p="No recent message available.";if(N.length>0){let F=[...N].sort((A,V)=>(V.createdAt||0)-(A.createdAt||0))[0],X=jJ({messages:[F],entities:O});if(X)p=X}let M=(F,X)=>{return F.sort((A,V)=>(A.createdAt||0)-(V.createdAt||0)).map((A)=>{let G=O.find((q)=>q.id===A.entityId)?.names[0]||(A.entityId===J.agentId?J.character.name:"Unknown"),j=A.createdAt?new Date(A.createdAt).toLocaleString():"Unknown time",L=A.content.text||"",z=X&&A.content.internalMonologue?`
214
+ [Internal thought: ${A.content.internalMonologue}]`:"";return`[${j}] ${G}: ${L}${z}`}).join(`
215
+ `)},w=M(N,!1),c=M(N,!0),h="";if(W){let F=`${W.messageCount} messages`,X=new Date(W.startTime).toLocaleDateString();if(h=`**Previous Conversation** (${F}, ${X})
216
+ `,h+=W.summary,W.topics&&W.topics.length>0)h+=`
217
+ *Topics: ${W.topics.join(", ")}*`;h=y("# Conversation Summary",h)}let QJ=Y.metadata,s=O.find((F)=>F.id===Y.entityId)?.names[0]||QJ?.entityName||"Unknown User",r=Y.content.text,n=!!r?.trim(),t=n?y("# Received Message",`${s}: ${r}`):"",d=n?y("# Focus your response",`You are replying to the above message from **${s}**. Keep your answer relevant to that message.`):"",a=new Map;if(B.length>0){let F=[...new Set(B.map((G)=>G.entityId).filter((G)=>G!==J.agentId))],X=new Set(F),A=new Set;O.forEach((G)=>{if(X.has(G.id))a.set(G.id,G),A.add(G.id)});let V=F.filter((G)=>!A.has(G));if(V.length>0)(await Promise.all(V.map((j)=>J.getEntityById(j)))).forEach((j,L)=>{if(j)a.set(V[L],j)})}let UJ=async(F)=>{return F.map((A)=>{let V=A.entityId===J.agentId,G;if(V)G=J.character.name;else G=a.get(A.entityId)?.metadata?.userName||"unknown";return`${G}: ${A.content.text}`}).join(`
218
+ `)},YJ=async(F,X)=>{let A=[...X],V=new Set(X.map((j)=>j.id));for(let[j,L]of a.entries())if(!V.has(j))A.push(L);return BJ({messages:F,entities:A,conversationHeader:!0})},[e,ZJ]=await Promise.all([UJ(B),YJ(B,O)]),FJ=[h,C,I,n?t:"",n?d:""].filter(Boolean).join(`
219
+
220
+ `);return{data:{summaries:W?[W]:[],recentMessages:N,recentInteractions:B,actionResults:l,mode:"summarized"},values:{...h&&{sessionSummaries:h},...C&&{recentMessages:C},recentMessageInteractions:e,recentPostInteractions:ZJ,recentInteractions:b?ZJ:e,...I&&{recentActionResults:I},recentMessage:p,conversationLog:w,conversationLogWithAgentThoughts:c,...t&&{receivedMessageHeader:t},...d&&{focusHeader:d}},text:FJ}}}catch(Z){return WK.error({error:Z},"Error in shortTermMemoryProvider:"),{data:{summaries:[],recentMessages:[],recentInteractions:[],actionResults:[],mode:"error"},values:{recentPosts:"",recentMessages:"",recentMessageInteractions:"",recentPostInteractions:"",recentInteractions:"",recentActionResults:"",conversationLog:"",conversationLogWithAgentThoughts:""},text:"Error retrieving conversation context."}}}};import{logger as AK,addHeader as GK}from"@elizaos/core";var pJ={name:"LONG_TERM_MEMORY",description:"Persistent facts and preferences about the user",position:50,get:async(J,Y,Q)=>{try{let Z=J.getService("memory");if(!Z)return{data:{memories:[]},values:{longTermMemories:""},text:""};let{entityId:K}=Y;if(K===J.agentId)return{data:{memories:[]},values:{longTermMemories:""},text:""};let _=await Z.getLongTermMemories(K,void 0,25);if(_.length===0)return{data:{memories:[]},values:{longTermMemories:""},text:""};let $=await Z.getFormattedLongTermMemories(K),U=GK("# What I Know About You",$),P=new Map;for(let R of _){let T=P.get(R.category)||0;P.set(R.category,T+1)}let W=Array.from(P.entries()).map(([R,T])=>`${R}: ${T}`).join(", ");return{data:{memories:_,categoryCounts:Object.fromEntries(P)},values:{longTermMemories:U,memoryCategories:W},text:U}}catch(Z){return AK.error({error:Z},"Error in longTermMemoryProvider:"),{data:{memories:[]},values:{longTermMemories:""},text:""}}}};var XK={name:"memory",description:"Advanced memory management with conversation summarization and long-term persistent memory",services:[KJ],evaluators:[TJ,SJ],providers:[pJ,IJ],schema:kJ},jK=XK;export{D as sessionSummaries,XK as memoryPlugin,bJ as memoryAccessLogs,H as longTermMemories,jK as default,KJ as MemoryService,EJ as LongTermMemoryCategory};
221
+
222
+ //# debugId=F6EA319C101EB9A964756E2164756E21