agent-working-memory 0.11.0 → 0.12.3

package/dist/mcp.js

@@ -6,13 +6,14 @@
  * Runs as a stdio-based MCP server that Claude Code connects to directly.
  * Uses the storage and engine layers in-process (no HTTP overhead).
  *
- * Tools exposed (16):
+ * Tools exposed (19):
  *   memory_write       — store a memory (salience filter decides disposition)
  *   memory_recall      — activate memories by context (cognitive retrieval)
  *   memory_feedback    — report whether a recalled memory was useful
  *   memory_retract     — invalidate a wrong memory with optional correction
  *   memory_supersede   — replace an outdated memory with a current one
  *   memory_stats       — get memory health metrics
+ *   memory_whoami      — identify this instance, mode, store, and sibling agent spaces
  *   memory_checkpoint  — save structured execution state (survives compaction)
  *   memory_restore     — restore state + targeted recall after compaction
  *   memory_task_add    — create a prioritized task
@@ -66,13 +67,19 @@ import { ConsolidationScheduler } from './engine/consolidation-scheduler.js';
 import { evaluateSalience } from './core/salience.js';
 import { performWrite } from './core/write-pipeline.js';
 import { DEFAULT_AGENT_CONFIG } from './types/agent.js';
-import { embed } from './core/embeddings.js';
+import { embed, getEmbedder } from './core/embeddings.js';
+import { getReranker } from './core/reranker.js';
+import { getExpander } from './core/query-expander.js';
 import { startSidecar } from './hooks/sidecar.js';
 import { initLogger, log, getLogPath } from './core/logger.js';
 import { VERSION } from './version.js';
 import { buildPack, INTERVIEW_QUESTIONS } from './onboard/index.js';
 import { liteCompress, retrieveOriginal } from './core/lite-compress.js';
 import { queryPeerDecisions, formatPeerDecisions } from './coordination/peer-decisions.js';
+import { startLoopLagMonitor } from './core/write-telemetry.js';
+import { buildWhoami, formatWhoami } from './core/whoami.js';
+import { renderTaskEndInvitation, validateRecipeWrite, recipeSlug, getRecipe } from './recipes/index.js';
+import { formatRecallResultLine } from './core/format-recall.js';
 // --- Incognito Mode ---
 // When AWM_INCOGNITO=1, register zero tools. Claude won't see memory tools at all.
 // No DB, no engines, no sidecar — just a bare MCP server that exposes nothing.
@@ -124,6 +131,7 @@ else {
     const consolidationScheduler = new ConsolidationScheduler(store, consolidationEngine);
     stagingBuffer.start(DEFAULT_AGENT_CONFIG.stagingTtlMs);
     consolidationScheduler.start();
+    startLoopLagMonitor();
     // Coordination DB handle — set when AWM_COORDINATION=true, used by memory_write for decision propagation
     let coordDb = null;
     const server = new McpServer({
@@ -188,15 +196,15 @@ else {
         return 'unclassified';
     }
     // --- Tools ---
-    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
-
-CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
-- Discover something about the codebase, bugs, or architecture
-- Make a decision and want to remember why
-- Encounter and resolve an error
-- Learn a user preference or project pattern
-- Complete a significant piece of work
-
+    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
+
+CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
+- Discover something about the codebase, bugs, or architecture
+- Make a decision and want to remember why
+- Encounter and resolve an error
+- Learn a user preference or project pattern
+- Complete a significant piece of work
+
 The concept should be a short label (3-8 words). The content should be the full detail.`, {
         concept: z.string().describe('Short label for this memory (3-8 words)'),
         content: z.string().describe('Full detail of what was learned'),
@@ -229,6 +237,14 @@ The concept should be a short label (3-8 words). The content should be the full
             .describe('Confidence: verified (tested), observed (read in code), assumed (reasoning).'),
         session_id: z.string().optional()
             .describe('Session/conversation grouping ID. Memories with same session_id are associated.'),
+        origin_class: z.enum(['user-stated', 'tool-output', 'inference', 'recipe']).optional()
+            .describe('Provenance (D5, log-only): where this knowledge came from. user-stated = the human said it; tool-output = read from a tool/system; inference = your reasoning; recipe = produced by a cognition recipe.'),
+        recipe_id: z.string().optional()
+            .describe('Cognition-recipe id+version when origin_class is recipe.'),
+        valid_from: z.string().optional()
+            .describe('ISO date when the FACT becomes valid (temporal validity, not ingestion time).'),
+        valid_to: z.string().optional()
+            .describe('ISO date when the FACT stops being valid (e.g., a deadline or a superseding change).'),
         intent: z.enum(['decision', 'question', 'todo', 'finding', 'context']).optional()
             .describe('What kind of memory this is.'),
     }, async (params) => {
@@ -247,6 +263,37 @@ The concept should be a short label (3-8 words). The content should be the full
             metaTags.push(`sid=${params.session_id}`);
         if (params.intent)
             metaTags.push(`intent=${params.intent}`);
+        // D14 (2026-07-30): recipe write-backs are contract-checked. Provenance
+        // must never claim a recipe that does not exist, and malformed
+        // derivations are rejected with the contract echoed back so the host
+        // can self-correct in one retry.
+        if (params.origin_class === 'recipe') {
+            if (!params.recipe_id) {
+                return { content: [{ type: 'text', text: "Recipe write rejected: origin_class 'recipe' requires recipe_id (e.g. 'skill-derivation@1')." }] };
+            }
+            const v = validateRecipeWrite(params.recipe_id, params.concept, params.content);
+            if (!v.ok) {
+                const contract = getRecipe(params.recipe_id)?.writeBack ?? 'unknown recipe';
+                return { content: [{ type: 'text', text: `Recipe write rejected (${params.recipe_id}): ${v.errors.join('; ')}\nContract: ${contract}` }] };
+            }
+            // Standardize recipe write-backs: canonical class, standard tags.
+            const slug = recipeSlug(params.concept);
+            const ensure = (tag) => { if (!userTags.includes(tag) && !metaTags.includes(tag))
+                metaTags.push(tag); };
+            if (params.recipe_id.startsWith('skill-derivation')) {
+                ensure('topic=skill');
+                ensure(`skill=${slug}`);
+                params.memory_type = params.memory_type ?? 'procedural';
+            }
+            else if (params.recipe_id.startsWith('friction-lesson')) {
+                ensure('topic=friction');
+                ensure(`about=${slug}`);
+                // zod defaults event_type to 'observation', so force the recipe's
+                // contract value rather than ??-guarding against undefined.
+                params.event_type = 'friction';
+            }
+            params.memory_class = 'canonical';
+        }
         const memoryType = params.memory_type ?? classifyMemoryType(params.content);
         const result = await performWrite({ store, connectionEngine }, {
             agentId: AGENT_ID,
@@ -261,6 +308,11 @@ The concept should be a short label (3-8 words). The content should be the full
             memoryClass: params.memory_class,
             memoryType,
             supersedes: params.supersedes,
+            originClass: params.origin_class,
+            writerSession: params.session_id,
+            recipeId: params.recipe_id,
+            validFrom: params.valid_from,
+            validTo: params.valid_to,
         });
         // Auto-checkpoint — covers create/reinforce/supersede uniformly
         try {
@@ -296,19 +348,22 @@ The concept should be a short label (3-8 words). The content should be the full
         return {
             content: [{
                     type: 'text',
-                    text: `Stored (${salience.disposition}) "${params.concept}" [${salience.score.toFixed(2)}]\nID: ${engram.id}`,
+                    text: `Stored (${salience.disposition}) "${params.concept}" [${salience.score.toFixed(2)}]\nID: ${engram.id}`
+                        + (isLowSalience
+                            ? `\nNOTE: low salience — this memory is kept but demoted and may fade first. If it MUST survive and be recallable, retry with memory_class: 'canonical'. (Discards are audited: reason codes ${JSON.stringify(salience.reasonCodes.slice(0, 4))})`
+                            : ''),
                 }],
         };
     });
-    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
-
-ALWAYS call this when:
-- Starting work on a project or topic (recall what you know)
-- Debugging (recall similar errors and solutions)
-- Making decisions (recall past decisions and outcomes)
-- The user mentions a topic you might have stored memories about
-
-Accepts either "query" or "context" parameter — both work identically.
+    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
+
+ALWAYS call this when:
+- Starting work on a project or topic (recall what you know)
+- Debugging (recall similar errors and solutions)
+- Making decisions (recall past decisions and outcomes)
+- The user mentions a topic you might have stored memories about
+
+Accepts either "query" or "context" parameter — both work identically.
 Returns the most relevant memories ranked by text relevance, temporal recency, and associative strength.`, {
         query: z.string().optional().describe('What to search for — describe the situation, question, or topic'),
         context: z.string().optional().describe('Alias for query (either works)'),
@@ -365,13 +420,11 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
                     }],
             };
         }
-        const lines = results.map((r, i) => {
-            // Confidence-adaptive output (Paper 3: cognitive teaming). When the caller
-            // requested 'compact' or 'auto' granularity, surface the engine-computed
-            // summary instead of the full content — same engram, less to read.
-            const body = r.summary ?? r.engram.content;
-            return `${i + 1}. **${r.engram.concept}** (${r.score.toFixed(3)}): ${body}`;
-        });
+        // Confidence-adaptive output (Paper 3: cognitive teaming) and D8
+        // (2026-07-30) conflict surfacing both live in the shared formatter now —
+        // see core/format-recall.ts for why it's extracted (0.12.1: unit-testable
+        // without booting the server) and why the id sits after the score.
+        const lines = results.map(formatRecallResultLine);
         return {
             content: [{
                     type: 'text',
@@ -379,8 +432,8 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
                 }],
         };
     });
-    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
-
+    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
+
 Always call this after using a recalled memory so the system learns what's valuable.`, {
         engram_id: z.string().describe('ID of the memory (from memory_recall results)'),
         useful: z.boolean().describe('Was this memory actually helpful?'),
@@ -403,8 +456,8 @@ Always call this after using a recalled memory so the system learns what's valua
                 }],
         };
     });
-    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
-
+    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
+
 Use this when you discover a memory contains incorrect information.`, {
         engram_id: z.string().describe('ID of the wrong memory'),
         reason: z.string().describe('Why is this memory wrong?'),
@@ -428,15 +481,15 @@ Use this when you discover a memory contains incorrect information.`, {
                 }],
         };
     });
-    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
-
-Use this when:
-- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
-- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
-- A schedule or plan was updated
-
+    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
+
+Use this when:
+- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
+- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
+- A schedule or plan was updated
+
 The old memory stays in the database (searchable for history) but is heavily down-ranked in recall so the current version dominates.`, {
-        old_engram_id: z.string().describe('ID of the outdated memory'),
+        old_engram_id: z.string().describe('ID of the outdated memory (from memory_recall results, or memory_write\'s own response if you just wrote it)'),
         new_engram_id: z.string().describe('ID of the replacement memory'),
         reason: z.string().optional().describe('Why the old memory is outdated'),
     }, async (params) => {
@@ -461,7 +514,11 @@ The old memory stays in the database (searchable for history) but is heavily dow
                 }],
         };
     });
-    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
+    server.tool('memory_whoami', `Identify THIS AWM instance — agent id, mode (standalone/hive), backend, store path, code provenance, ports, and the sibling agent spaces present in the same store. Call when unsure which AWM instance or memory space you are talking to.`, {}, async () => {
+        const info = await buildWhoami(store, AGENT_ID, 'mcp');
+        return { content: [{ type: 'text', text: formatWhoami(info) }] };
+    });
+    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
 Also shows the activity log path so the user can tail it to see what's happening.`, {}, async () => {
         const metrics = await evalEngine.computeMetrics(AGENT_ID);
         const checkpoint = await store.getCheckpoint(AGENT_ID);
@@ -492,13 +549,13 @@ Also shows the activity log path so the user can tail it to see what's happening
         };
     });
     // --- Checkpointing Tools ---
-    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
-
-ALWAYS call this before:
-- Long operations (multi-file generation, large refactors, overnight work)
-- Anything that might fill the context window
-- Switching to a different task
-
+    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
+
+ALWAYS call this before:
+- Long operations (multi-file generation, large refactors, overnight work)
+- Anything that might fill the context window
+- Switching to a different task
+
 Also call periodically during long sessions to avoid losing state. The state is saved per-agent and overwrites any previous checkpoint.`, {
         current_task: z.string().describe('What you are currently working on'),
         decisions: z.array(z.string()).optional().default([])
@@ -532,14 +589,14 @@ Also call periodically during long sessions to avoid losing state. The state is
                 }],
         };
     });
-    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
-
-Returns:
-- Your saved execution state (task, decisions, next steps, files)
-- Recently recalled memories for context
-- Your last write for continuity
-- How long you were idle
-
+    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
+
+Returns:
+- Your saved execution state (task, decisions, next steps, files)
+- Recently recalled memories for context
+- Your last write for continuity
+- How long you were idle
+
 Use this at the start of every session or after compaction to pick up where you left off.`, {}, async () => {
         const checkpoint = await store.getCheckpoint(AGENT_ID);
         // Cold-store nudge: an empty store means the agent has nothing to recall — offer to warm-start.
@@ -658,9 +715,9 @@ Use this at the start of every session or after compaction to pick up where you
         if (coordDb) {
             try {
                 const myAgent = coordDb.prepare(`SELECT id FROM coord_agents WHERE name = ? AND status != 'dead' ORDER BY last_seen DESC LIMIT 1`).get(AGENT_ID);
-                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
-           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
-           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
+                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
+           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
+           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
            ORDER BY d.created_at DESC LIMIT 10`).all(myAgent?.id ?? '');
                 if (peerDecisions.length > 0) {
                     parts.push(`\n**Peer decisions (last 30 min):**`);
@@ -679,11 +736,11 @@ Use this at the start of every session or after compaction to pick up where you
         };
     });
     // --- Onboarding Tools (warm-start a cold store) ---
-    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
-
-Use this when the store is empty / you're new to a project. The scan is deterministic
-(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
-recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
+    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
+
+Use this when the store is empty / you're new to a project. The scan is deterministic
+(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
+recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
 save the good ones with memory_write (memory_class="canonical"). Nothing is saved by this tool.`, {
         docs: z.array(z.string()).optional()
             .describe('Doc files/dirs to scan (Markdown/text). Defaults to the repo (or cwd).'),
@@ -708,18 +765,18 @@ save the good ones with memory_write (memory_class="canonical"). Nothing is save
         ].join('\n');
         return { content: [{ type: 'text', text }] };
     });
-    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
+    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
 goal of the memory system, and ask follow-ups for clarity. Turn each answer into a canonical memory.`, {}, async () => ({
         content: [{ type: 'text', text: INTERVIEW_QUESTIONS.map((q, i) => `${i + 1}. ${q}`).join('\n') }],
     }));
     // --- Task Management Tools ---
-    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
-
-Use this when:
-- You identify work that needs doing but can't do it right now
-- The user mentions something to do later
-- You want to park a sub-task while focusing on something more urgent
-
+    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
+
+Use this when:
+- You identify work that needs doing but can't do it right now
+- The user mentions something to do later
+- You want to park a sub-task while focusing on something more urgent
+
 Tasks automatically get high salience so they won't be discarded.`, {
         concept: z.string().describe('Short task title (3-10 words)'),
         content: z.string().describe('Full task description — what needs doing, context, acceptance criteria'),
@@ -759,11 +816,11 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
-- Start working on a task (open → in_progress)
-- Mark a task done (→ done)
-- Block a task on another (→ blocked)
-- Reprioritize (change priority)
+    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
+- Start working on a task (open → in_progress)
+- Mark a task done (→ done)
+- Block a task on another (→ blocked)
+- Reprioritize (change priority)
 - Unblock a task (clear blocked_by)`, {
         task_id: z.string().describe('ID of the task to update'),
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
@@ -793,8 +850,8 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
-
+    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
+
 Use at the start of a session to see what's pending, or to check blocked/done tasks.`, {
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
             .describe('Filter by status (omit to see all active tasks)'),
@@ -820,10 +877,10 @@ Use at the start of a session to see what's pending, or to check blocked/done ta
                 }],
         };
     });
-    server.tool('memory_task_next', `Get the single most important task to work on next.
-
-Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
-
+    server.tool('memory_task_next', `Get the single most important task to work on next.
+
+Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
+
 Use this when you finish a task or need to decide what to do next.`, {}, async () => {
         const next = await store.getNextTask(AGENT_ID);
         if (!next) {
@@ -839,13 +896,13 @@ Use this when you finish a task or need to decide what to do next.`, {}, async (
         };
     });
     // --- Task Bracket Tools ---
-    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
-
-CALL THIS when starting:
-- A multi-step operation (doc generation, large refactor, migration)
-- Work on a new topic or project area
-- Anything that might fill the context window
-
+    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
+
+CALL THIS when starting:
+- A multi-step operation (doc generation, large refactor, migration)
+- Work on a new topic or project area
+- Anything that might fill the context window
+
 This ensures your state is saved before you start, and primes recall with relevant context.`, {
         topic: z.string().describe('What task are you starting? (3-15 words)'),
         files: z.array(z.string()).optional().default([])
@@ -896,13 +953,13 @@ This ensures your state is saved before you start, and primes recall with releva
                 }],
         };
     });
-    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
-
-CALL THIS when you finish:
-- A multi-step operation
-- Before switching to a different topic
-- At the end of a work session
-
+    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
+
+CALL THIS when you finish:
+- A multi-step operation
+- Before switching to a different topic
+- At the end of a work session
+
 This captures what was accomplished so future sessions can recall it.`, {
         summary: z.string().describe('What was accomplished? Include key outcomes, decisions, and any issues.'),
         tags: z.array(z.string()).optional().default([])
@@ -971,18 +1028,21 @@ This captures what was accomplished so future sessions can recall it.`, {
         return {
             content: [{
                     type: 'text',
-                    text: `Completed: "${completedTask}" [${salience.score.toFixed(2)}]${supersededNote}`,
+                    // D14 (2026-07-30): task end is the recipe moment — invite the host
+                    // to distill a skill and/or a failure lesson in separate focused
+                    // passes. The host owns the gates; AWM validates the write-backs.
+                    text: `Completed: "${completedTask}" [${salience.score.toFixed(2)}]${supersededNote}\n${renderTaskEndInvitation()}`,
                 }],
         };
     });
-    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
-a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
-of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
-accurately as JSON). Use this on large tool results you need to keep in context.
-
-Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
-TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
-otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
+    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
+a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
+of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
+accurately as JSON). Use this on large tool results you need to keep in context.
+
+Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
+TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
+otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
 retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
         output: z.string().describe('The tool output to compress — JSON text (preferred) or any string. Non-JSON is returned unchanged.'),
         min_saving_chars: z.number().optional().describe('Only emit TOON if it saves at least this many characters (default 40).'),
@@ -996,8 +1056,8 @@ retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
             content: [{ type: 'text', text: header + r.text }],
         };
     });
-    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
-you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
+    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
+you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
 an error if the ref has expired (originals are kept for the most recent compressions only).`, {
         ref: z.string().describe('The ref handle returned by compress_output (e.g. "awm_orig_12").'),
     }, async (params) => {
@@ -1021,6 +1081,35 @@ an error if the ref has expired (originals are kept for the most recent compress
             agentId: AGENT_ID,
             secret: HOOK_SECRET,
             port: HOOK_PORT,
+            // 0.12.2: warm recall for hooks — the sidecar shares this process's
+            // activation engine and loaded models, so a UserPromptSubmit hook can get
+            // warm-latency recall without any standing server. Trimmed result shape
+            // (no embeddings/phase scores — hooks don't need them and the vectors
+            // alone would 10x the payload).
+            activate: async (q) => {
+                const results = await activationEngine.activate({
+                    agentId: AGENT_ID,
+                    context: q.context,
+                    limit: q.limit,
+                    requireConfidence: q.requireConfidence,
+                    granularity: q.granularity,
+                });
+                return results.map(r => ({
+                    engram: {
+                        id: r.engram.id,
+                        concept: r.engram.concept,
+                        content: r.engram.content,
+                        createdAt: r.engram.createdAt instanceof Date
+                            ? r.engram.createdAt.toISOString()
+                            : r.engram.createdAt,
+                        memoryClass: r.engram.memoryClass,
+                        validTo: r.engram.validTo,
+                    },
+                    score: r.score,
+                    summary: r.summary,
+                    confidence: r.confidence,
+                }));
+            },
             onConsolidate: async (agentId, reason) => {
                 console.error(`[mcp] consolidation triggered: ${reason}`);
                 const result = await consolidationEngine.consolidate(agentId);
@@ -1028,6 +1117,31 @@ an error if the ref has expired (originals are kept for the most recent compress
                 console.error(`[mcp] consolidation done: ${result.edgesStrengthened} strengthened, ${result.memoriesForgotten} forgotten`);
             },
         });
+        // 0.12.2: eager warm — fire-and-forget, mirrors index.ts:208-218. Without
+        // this, every Claude Code session paid the full cold cost (~3s measured on
+        // a 29.7k-engram store: slim cache ~0.9s + three model loads ~1.8s) on its
+        // FIRST recall, which is exactly the "first recall is slow → recall gets
+        // avoided" failure mode. Warming here overlaps session startup instead.
+        // All output is stderr-safe (stdout carries JSON-RPC frames).
+        // Escape hatch: AWM_NO_EAGER_WARM=1 restores lazy loading.
+        if (process.env.AWM_NO_EAGER_WARM !== '1') {
+            getEmbedder().catch(err => console.error('Embedding model unavailable:', err.message));
+            getReranker().catch(err => console.error('Reranker model unavailable:', err.message));
+            getExpander().catch(err => console.error('Query expander model unavailable:', err.message));
+            if (BACKEND === 'sqlite') {
+                setImmediate(() => {
+                    try {
+                        const t0 = Date.now();
+                        store.warmSlimCache();
+                        const stats = store.getSlimCacheStats();
+                        console.error(`Slim cache warmed: ${stats.size} entries in ${Date.now() - t0}ms`);
+                    }
+                    catch (err) {
+                        console.error(`Slim cache warm failed: ${err.message}`);
+                    }
+                });
+            }
+        }
         // Coordination MCP tools (opt-in via AWM_COORDINATION=true)
         // AWM 0.8.x: coordination requires SQLite (uses store.getDb()). On PGlite,
         // coordination is auto-disabled with a warning; re-enable when coordination