th-memory-mcp 2.0.0 → 2.1.0

package/ARCHITECTURE_v2.md

@@ -0,0 +1,1582 @@
+# th-memory-mcp v2 — Architecture & Implementation Specification
+
+**Status:** ✅ Released — `th-memory-mcp v2.0.0` is published (npm + Official MCP Registry + Glama).  
+**Baseline:** v1.2.2 → **Current:** v2.0.0  
+**Primary goal:** evolve th-memory-mcp from a structured local memory MCP into a durable, temporal, conflict-aware, hybrid-retrieval memory engine for AI agents.
+
+> **Audience guide:** End users should read [README.md](README.md) (install, tools, usage). This document is the **canonical architecture & agent-rules spec** for developers and AI coding agents — the single source of truth for structure and behavior. The former `design.md` build log has been folded into §40 Implementation Status.
+
+---
+
+## 0. Executive Decision
+
+v2 is a **refactor + controlled expansion**, not a rewrite and not a clone of Mem0, Zep/Graphiti, or Letta.
+
+The project must retain these v1 properties:
+
+- Local-first and offline by default.
+- SQLite as the primary persistence layer.
+- FTS5 and local semantic search.
+- Thai/English support.
+- Auto-capture and cross-harness compatibility.
+- Secret filtering and safe export.
+- Graceful degradation when memory is unavailable.
+- Small MCP surface and bounded output.
+
+v2 adds five core capabilities:
+
+1. Unified memory model.
+2. Temporal state and supersession.
+3. Duplicate/conflict resolution.
+4. Hybrid retrieval with ranking/fusion.
+5. Context assembly with token budgeting.
+
+Optional AI-assisted extraction/consolidation must never make the core memory engine dependent on an external LLM API.
+
+---
+
+# 1. Design Principles
+
+## 1.1 Memory is data, not instructions
+
+Stored memory must never override the agent's system/developer instructions or become executable instructions merely because it contains imperative text.
+
+## 1.2 Event != memory
+
+Raw interactions are evidence/feedstock. Long-term memories are derived, structured records.
+
+```text
+Interaction/Event
+      ↓
+Capture + filtering
+      ↓
+Extraction/classification
+      ↓
+Memory candidate
+      ↓
+Dedup/conflict resolution
+      ↓
+Persistent memory
+```
+
+## 1.3 Current truth and historical truth are both valuable
+
+Old memories are not automatically deleted because they became stale. They can remain available for historical queries.
+
+## 1.4 Deterministic-first
+
+Core operations must work without an LLM:
+
+- persistence
+- FTS search
+- semantic search
+- metadata filtering
+- scoring
+- RRF fusion
+- lifecycle transitions
+- basic duplicate detection
+- basic conflict detection
+
+LLM assistance is optional for:
+
+- difficult extraction
+- ambiguous conflict resolution
+- consolidation
+- summarization/compression
+
+## 1.5 Context is a projection of memory
+
+The database is not the prompt. `get_context` selects a small, relevant, safe projection from persistent memory.
+
+## 1.6 Failure must be non-fatal
+
+Memory failures must not crash the host agent. MCP operations should return bounded diagnostic text where appropriate and continue gracefully.
+
+---
+
+# 2. High-Level Architecture
+
+```text
+                         AI AGENT
+                            │
+                           MCP
+                            │
+                    ┌───────▼────────┐
+                    │  MEMORY API    │
+                    └───────┬────────┘
+                            │
+              ┌─────────────┼─────────────┐
+              │             │             │
+              ▼             ▼             ▼
+        CAPTURE ENGINE  RETRIEVAL      LIFECYCLE
+              │          ENGINE         ENGINE
+              │             │             │
+              │      ┌──────┼──────┐      │
+              │      │      │      │      │
+              │     FTS   VECTOR  GRAPH  │
+              │      │      │      │      │
+              │      └──────┼──────┘      │
+              │             │             │
+              └─────────────┼─────────────┘
+                            ▼
+                     MEMORY STORE
+                ┌─────────────────────┐
+                │ SQLite              │
+                │ FTS5                │
+                │ local vectors       │
+                │ entities/relations  │
+                │ temporal metadata   │
+                └──────────┬──────────┘
+                           │
+                     CONTEXT ENGINE
+                           │
+                           ▼
+                        AI AGENT
+```
+
+---
+
+# 3. Memory Taxonomy
+
+Canonical memory types:
+
+```text
+FACT
+PREFERENCE
+GOAL
+DECISION
+CONSTRAINT
+LESSON
+PROCEDURE
+EPISODE
+RELATION
+PROFILE
+```
+
+### Semantics
+
+| Type | Purpose |
+|---|---|
+| FACT | durable factual information |
+| PREFERENCE | user/project preference |
+| GOAL | desired future outcome |
+| DECISION | chosen approach and rationale |
+| CONSTRAINT | hard requirement or prohibition |
+| LESSON | correction-derived knowledge |
+| PROCEDURE | reusable method/workflow |
+| EPISODE | meaningful historical event |
+| RELATION | entity relationship information |
+| PROFILE | high-value compact user/project summary |
+
+Types must be extensible internally, but these ten are the stable v2 vocabulary.
+
+---
+
+# 4. Memory Lifecycle
+
+Every memory has a lifecycle state:
+
+```text
+NEW → ACTIVE
+ACTIVE → REINFORCED → ACTIVE
+ACTIVE → STALE
+ACTIVE → SUPERSEDED
+STALE → ARCHIVED
+SUPERSEDED → ARCHIVED
+ACTIVE → DELETED
+ARCHIVED → DELETED
+```
+
+### Rules
+
+- `ACTIVE`: eligible for normal retrieval.
+- `STALE`: low priority; eligible when historical context is useful.
+- `SUPERSEDED`: replaced by another memory; normally excluded from current-context retrieval.
+- `ARCHIVED`: retained but excluded from default retrieval.
+- `DELETED`: logically deleted unless hard-delete is explicitly requested.
+
+A superseded memory should retain a pointer to the replacement.
+
+---
+
+# 5. Temporal Model
+
+Each memory may have both record time and validity time:
+
+```text
+created_at
+updated_at
+last_accessed_at
+valid_from
+valid_until
+```
+
+`valid_until = NULL` means currently valid unless lifecycle state says otherwise.
+
+Temporal questions must be supported conceptually:
+
+- What is true now?
+- What was true at time T?
+- What changed?
+- Which memory superseded this one?
+
+Do not physically delete historical truth merely because it is no longer current.
+
+---
+
+# 6. Database Schema
+
+The following is the logical v2 schema. Migration SQL may implement equivalent SQLite details, but semantics must remain compatible.
+
+## 6.1 `memories`
+
+```sql
+CREATE TABLE memories (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  type TEXT NOT NULL,
+  content TEXT NOT NULL,
+  summary TEXT,
+  status TEXT NOT NULL DEFAULT 'active',
+  source TEXT NOT NULL DEFAULT 'explicit',
+  confidence REAL NOT NULL DEFAULT 0.5,
+  importance REAL NOT NULL DEFAULT 0.5,
+  salience REAL NOT NULL DEFAULT 0.5,
+  project_id TEXT,
+  session_id TEXT,
+  created_at TEXT NOT NULL,
+  updated_at TEXT NOT NULL,
+  last_accessed_at TEXT,
+  access_count INTEGER NOT NULL DEFAULT 0,
+  valid_from TEXT,
+  valid_until TEXT,
+  supersedes_id INTEGER,
+  metadata TEXT,
+  FOREIGN KEY (supersedes_id) REFERENCES memories(id)
+);
+```
+
+### Required indexes
+
+```sql
+CREATE INDEX idx_memories_type_status ON memories(type, status);
+CREATE INDEX idx_memories_project_status ON memories(project_id, status);
+CREATE INDEX idx_memories_updated ON memories(updated_at);
+CREATE INDEX idx_memories_validity ON memories(valid_from, valid_until);
+CREATE INDEX idx_memories_supersedes ON memories(supersedes_id);
+```
+
+## 6.2 `interactions`
+
+Retain raw behavior/event storage:
+
+```sql
+CREATE TABLE interactions (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  ts TEXT NOT NULL,
+  session_id TEXT,
+  kind TEXT NOT NULL,
+  content TEXT NOT NULL,
+  meta TEXT
+);
+```
+
+Interactions are not automatically long-term memories.
+
+## 6.3 `entities`
+
+```sql
+CREATE TABLE entities (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  name TEXT NOT NULL,
+  canonical_name TEXT NOT NULL,
+  type TEXT,
+  metadata TEXT
+);
+CREATE UNIQUE INDEX idx_entities_canonical ON entities(canonical_name);
+```
+
+## 6.4 `relations`
+
+```sql
+CREATE TABLE relations (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  source_entity_id INTEGER NOT NULL,
+  relation TEXT NOT NULL,
+  target_entity_id INTEGER NOT NULL,
+  confidence REAL NOT NULL DEFAULT 0.5,
+  valid_from TEXT,
+  valid_until TEXT,
+  source_memory_id INTEGER,
+  metadata TEXT,
+  FOREIGN KEY (source_entity_id) REFERENCES entities(id),
+  FOREIGN KEY (target_entity_id) REFERENCES entities(id),
+  FOREIGN KEY (source_memory_id) REFERENCES memories(id)
+);
+```
+
+## 6.5 `memory_links`
+
+```sql
+CREATE TABLE memory_links (
+  source_memory_id INTEGER NOT NULL,
+  relation TEXT NOT NULL,
+  target_memory_id INTEGER NOT NULL,
+  confidence REAL NOT NULL DEFAULT 0.5,
+  created_at TEXT NOT NULL,
+  PRIMARY KEY (source_memory_id, relation, target_memory_id),
+  FOREIGN KEY (source_memory_id) REFERENCES memories(id),
+  FOREIGN KEY (target_memory_id) REFERENCES memories(id)
+);
+```
+
+Supported link relations:
+
+```text
+supports
+contradicts
+supersedes
+derived_from
+related_to
+caused_by
+depends_on
+```
+
+## 6.6 `profile`
+
+Retain compact profile sections for backward compatibility and fast injection:
+
+```sql
+CREATE TABLE profile (
+  section TEXT PRIMARY KEY,
+  content TEXT NOT NULL,
+  updated_at TEXT NOT NULL
+);
+```
+
+In v2, profile is a projection/cache of important memories, not the canonical source of truth.
+
+---
+
+# 7. Search Indexes
+
+FTS5 remains mandatory.
+
+Logical indexed fields:
+
+```text
+memory id
+memory type
+content
+summary
+project_id
+```
+
+The implementation may use a maintained FTS5 virtual table or separate indexes, but every mutation of searchable memory must keep indexes synchronized transactionally where possible.
+
+Local semantic search remains supported. The implementation must preserve the dependency-light/offline property of v1.
+
+---
+
+# 8. Memory Source Model
+
+Canonical source values:
+
+```text
+explicit
+corrected
+inferred
+captured
+consolidated
+imported
+system
+```
+
+Recommended source weights for confidence calculation:
+
+| Source | Weight |
+|---|---:|
+| explicit | 1.00 |
+| corrected | 0.95 |
+| captured | 0.30 |
+| inferred | 0.50 |
+| consolidated | 0.75 |
+| imported | 0.70 |
+| system | 0.80 |
+
+These are defaults, not immutable constants.
+
+---
+
+# 9. Confidence, Importance, Salience
+
+All three values are normalized to `[0,1]`.
+
+## 9.1 Confidence
+
+Confidence reflects how trustworthy the memory is.
+
+Recommended conceptual model:
+
+```text
+confidence = f(source_weight,
+               confirmation_count,
+               consistency,
+               conflict_penalty)
+```
+
+Do not blindly increase confidence forever from duplicate saves. Repeated identical events should have diminishing returns.
+
+## 9.2 Importance
+
+Importance is durable significance. It should not decay merely because the memory is old.
+
+Examples:
+
+- hard project constraint: high
+- architectural decision: high
+- temporary debugging detail: low
+
+## 9.3 Salience
+
+Salience determines usefulness for a particular retrieval/context operation.
+
+Recommended baseline:
+
+```text
+salience =
+  0.30 * semantic_relevance +
+  0.20 * importance +
+  0.15 * confidence +
+  0.15 * recency +
+  0.10 * access_frequency +
+  0.10 * project_relevance
+```
+
+Weights must be configurable and benchmarked.
+
+---
+
+# 10. Recency and Decay
+
+Use a bounded exponential recency factor:
+
+```text
+recency = exp(-lambda * age_days)
+```
+
+Decay policy must depend on memory type.
+
+| Type | Default decay |
+|---|---|
+| CONSTRAINT | very low |
+| DECISION | low |
+| LESSON | low |
+| PREFERENCE | low |
+| FACT | low/medium |
+| GOAL | medium |
+| PROCEDURE | low/medium |
+| EPISODE | medium |
+| RELATION | low/medium |
+| PROFILE | derived |
+
+These are policy classes, not fixed numeric constants.
+
+---
+
+# 11. Deduplication
+
+Every `remember` candidate must pass duplicate detection before insertion.
+
+Pipeline:
+
+```text
+normalize
+  ↓
+exact match
+  ↓
+canonical/key match
+  ↓
+FTS similarity
+  ↓
+semantic similarity (if available)
+  ↓
+DUPLICATE / UPDATE / DISTINCT
+```
+
+Duplicate detection must avoid merging memories that are merely similar but semantically different.
+
+---
+
+# 12. Conflict Resolution
+
+This is a first-class v2 subsystem.
+
+When a new candidate arrives:
+
+```text
+candidate
+   ↓
+find related active memories
+   ↓
+classify relationship
+   ├── duplicate
+   ├── update
+   ├── contradiction
+   └── unrelated
+```
+
+### Update/supersession
+
+For a direct change:
+
+```text
+old memory: status = superseded
+new memory: status = active
+new memory.supersedes_id = old.id
+```
+
+Create a `supersedes` memory link when useful.
+
+### Ambiguous conflict
+
+If deterministic rules cannot safely decide, preserve both records and mark the relationship `contradicts`; do not silently destroy information.
+
+Optional AI-assisted resolution may later select the current truth based on explicit user evidence.
+
+---
+
+# 13. Hybrid Retrieval
+
+`recall` and `get_context` must use more than one retrieval signal.
+
+```text
+QUERY
+ │
+ ├── FTS5 keyword retrieval
+ │
+ ├── local semantic retrieval
+ │
+ ├── metadata/scope filtering
+ │
+ └── optional graph expansion
+ │
+ ▼
+Candidate pool
+ │
+ ▼
+RRF fusion
+ │
+▼
+Scoring/reranking
+ │
+▼
+Conflict/status filtering
+ │
+▼
+Top K
+```
+
+## 13.1 Reciprocal Rank Fusion
+
+Use RRF rather than directly mixing incompatible raw search scores:
+
+```text
+RRF(m) = Σ 1 / (k + rank_i(m))
+```
+
+Then apply memory-specific factors:
+
+```text
+final_score =
+  RRF * confidence * importance_factor * recency_factor * scope_factor
+```
+
+The exact formula must be benchmarked.
+
+---
+
+# 14. Graph Retrieval
+
+Graph is an augmentation, not the sole retrieval mechanism.
+
+Default traversal should be shallow and bounded:
+
+```text
+seed entities/memories
+      ↓
+1-hop related entities
+      ↓
+related memories
+```
+
+Do not perform unbounded graph traversal during normal `recall`.
+
+Graph boost should improve relationship queries without allowing weak graph edges to dominate strong direct evidence.
+
+---
+
+# 15. Context Engine
+
+Introduce a first-class `get_context` operation.
+
+Input concept:
+
+```json
+{
+  "query": "current task",
+  "project": "optional-project-id",
+  "limit": 12,
+  "token_budget": 1500,
+  "include_history": false
+}
+```
+
+Pipeline:
+
+```text
+query
+ ↓
+retrieve candidate memories
+ ↓
+filter stale/superseded records
+ ↓
+deduplicate
+ ↓
+resolve conflicts
+ ↓
+rank
+ ↓
+fit token budget
+ ↓
+assemble structured context
+```
+
+Output should be concise and machine-readable enough for agents to consume.
+
+Recommended sections:
+
+```text
+Current Profile
+Relevant Preferences
+Constraints
+Decisions
+Lessons
+Relevant Facts
+Historical Context (only when requested/useful)
+```
+
+Default context should prioritize current project + active user constraints.
+
+---
+
+# 16. Persistent vs Archival Memory
+
+Two logical tiers:
+
+## Tier A — Persistent/Pinned
+
+Small, high-value context that may be injected without retrieval:
+
+- identity/profile essentials
+- active goals
+- critical constraints
+- current project decisions
+
+Target size: approximately 500–1500 tokens depending on host/context budget.
+
+## Tier B — Archival
+
+Searchable persistent memory:
+
+- facts
+- episodes
+- lessons
+- procedures
+- historical decisions
+- relations
+
+Tier A is a projection/cache; Tier B remains the source of truth.
+
+---
+
+# 17. MCP Tool Surface
+
+Do not expand to dozens of tools. v2.1.0 ships **16 tools** (the original spec targeted 14; the extra 2 are `consolidate` and `extract_memories`, which extend the v2 engine).
+
+## Shipped (16 tools)
+
+### Core / compatibility (carried from v1)
+1. `remember` — store a memory using the unified model (type + metadata)
+2. `recall` — hybrid FTS + semantic search (v1 behavior preserved)
+3. `forget` — soft delete by default
+4. `get_profile` — compact profile projection/cache
+5. `search_history` — search raw interactions
+6. `get_recent_interactions` — recent raw events
+7. `memory_stats` — lifecycle/retrieval/storage metrics
+8. `export_memory` — safe export confined to the allowed directory
+9. `save_lesson` — compatibility alias/wrapper over `remember(type=LESSON)`
+
+### v2 engine tools
+10. `get_context` — token-budgeted context assembly (preferred agent-facing retrieval)
+11. `consolidate` — cluster related memories + optional derived memory
+12. `link_memory` — typed relationship between two memories in the graph
+13. `merge_memory` — merge a duplicate into a canonical memory (source superseded, provenance in `metadata.merged_from`)
+14. `update_memory` — update mutable fields in place, or create a superseding memory when `content` changes
+15. `import_memory` — import memories from JSON (validates type, dedupes, dry-run by default)
+16. `extract_memories` — scan recent interactions for memory-intent and propose/create memories (deterministic, no LLM; dry-run by default)
+
+All 16 are documented in README.md. The four tools originally marked "deferred" in this spec (`update_memory`, `merge_memory`, `link_memory`, `import_memory`) plus `extract_memories` were implemented in v2.1.0.
+
+---
+
+# 18. Tool Contracts
+
+## `remember`
+
+```text
+content: string
+ type?: MemoryType
+importance?: 0..1
+confidence?: 0..1
+project_id?: string
+session_id?: string
+source?: SourceType
+valid_from?: ISO timestamp
+valid_until?: ISO timestamp
+metadata?: object
+```
+
+Returns:
+
+```text
+created | reinforced | updated | superseded | duplicate | conflict
+memory id
+short summary
+```
+
+## `recall`
+
+```text
+query: string
+limit?: 1..50
+project_id?: string
+include_archived?: boolean
+include_history?: boolean
+```
+
+Uses hybrid retrieval.
+
+## `get_context`
+
+```text
+query?: string
+project_id?: string
+token_budget?: integer
+limit?: integer
+include_history?: boolean
+```
+
+This should be the preferred agent-facing retrieval operation for complex tasks.
+
+## `update_memory`
+
+Updates mutable fields without silently changing identity/history.
+
+Changes to factual content that represent a new truth should create a superseding memory where appropriate. (Implemented in v2.1.0 — see `src/tools/update_memory.ts`.)
+
+## `merge_memory`
+
+Combines duplicate/near-duplicate memories while preserving provenance and source IDs. (Implemented in v2.1.0 — see `src/tools/merge_memory.ts`.)
+
+## `link_memory`
+
+Creates a typed relationship between memories. (Implemented in v2.1.0 — see `src/tools/link_memory.ts`.)
+
+## `extract_memories`
+
+Scans recent captured interactions for memory-intent phrases and proposes memory candidates. Deterministic (no LLM); dry-run by default, `apply=true` creates them (source=`captured`). Implemented in v2.1.0 — see `src/tools/extract_memories.ts`.
+
+## `forget`
+
+Default behavior is soft delete. Hard delete must be explicit and documented.
+
+## `consolidate_memory`
+
+Groups related memories and creates compact derived memories without automatically deleting source evidence.
+
+## `memory_stats`
+
+Expose lifecycle, retrieval, quality, and storage metrics.
+
+## `export_memory` / `import_memory`
+
+Must preserve IDs only when safe. Imports must validate schema/version and never overwrite active memory blindly.
+
+---
+
+# 19. Auto-Capture v2
+
+Retain the existing plugin architecture but make capture a pipeline.
+
+```text
+Host event
+ ↓
+Secret/PII filtering
+ ↓
+Noise filter
+ ↓
+Deduplication
+ ↓
+Interaction event storage
+ ↓
+Optional extraction
+ ↓
+Memory candidate
+```
+
+Capture must never turn every user prompt into long-term memory.
+
+### Existing v1 safeguards to retain
+
+- secret filtering
+- event deduplication
+- truncation
+- try/catch around every write
+- shared `MEMORY_DB_PATH`
+- WAL
+
+---
+
+# 20. Security Requirements
+
+Required:
+
+- secret filtering before persistence
+- safe export filenames
+- export confined to allowed export directory
+- metadata sanitization
+- memory treated as untrusted data
+- no memory-generated command execution
+- no memory-generated system/developer instruction override
+- no remote service required by default
+
+Sensitive data must not be reintroduced merely because it was previously stored. Existing v1 filtering rules are the minimum baseline, not the maximum security boundary.
+
+---
+
+# 21. Graceful Degradation
+
+If any subsystem fails:
+
+```text
+Graph unavailable → use FTS/vector
+Vector unavailable → use FTS
+FTS unavailable → use basic SQL filtering
+Profile unavailable → continue without profile
+DB unavailable → return bounded error and allow agent to continue
+```
+
+No optional subsystem should become a single point of failure for the MCP server.
+
+---
+
+# 22. Source Tree (as-built in v2.0.0)
+
+```text
+src/
+├── index.ts                     # MCP server: registers 11 tools
+├── db/
+│   ├── index.ts                 # db singleton (better-sqlite3, WAL) + runMigrations
+│   ├── migrations.ts            # ordered MIGRATIONS array (TS modules, idempotent)
+│   └── repositories/
+│       └── memories.ts          # CRUD + searchMemories (delegates to hybrid retrieval)
+├── memory/
+│   ├── types.ts                 # MemoryType, SourceType, LifecycleState, Scope, MemoryRecord
+│   ├── decay.ts                 # recencyFactor + per-type DECAY_LAMBDA policy classes
+│   ├── source-weights.ts        # SOURCE_WEIGHTS
+│   ├── scorer.ts                # computeSalience / computeConfidence
+│   ├── deduplicator.ts          # normalize / exact / similar / deduplicate
+│   └── conflict-resolver.ts     # isContradiction / classifyRelationship / resolveConflict
+├── retrieval/
+│   ├── fts.ts                    # FTS5 search
+│   ├── vector.ts                 # cosine over embeddings
+│   ├── fusion.ts                 # rrfFuse (k=60)
+│   └── scorer.ts                # finalScore (RRF × confidence × importance × recency × scope)
+├── core/
+│   ├── lifecycle-engine.ts       # transitions, reinforce, supersede, archive, softDelete
+│   ├── temporal-engine.ts        # validity intervals, point-in-time, supersession chains
+│   ├── retrieval-engine.ts       # retrieve(): FTS+vector → RRF → score → filter → topK
+│   ├── graph-engine.ts           # entities/relations, linkMemories, bounded traverse
+│   ├── context-engine.ts         # getContext(): retrieve → graph expand → temporal filter → budget
+│   └── consolidation-engine.ts   # clusterMemories, createDerivedMemory, getProvenance
+├── tools/
+│   ├── context.ts                # get_context tool
+│   └── consolidate.ts            # consolidate tool
+├── lib/
+│   └── embed.ts                  # hashing-trick embed + DataView serialize/deserialize (fixed in v2)
+└── plugin/
+    └── learning-capture.ts        # Bun auto-capture plugin (OpenCode)
+scripts/
+└── claude-capture.mjs            # Claude Code hook capture
+test/
+└── *.test.mjs                     # 12 suites (capture, distill, lifecycle, temporal, conflict,
+                                   #   retrieval, graph, context, consolidation, benchmark, security, smoke)
+```
+
+Compatibility wrappers keep the old `remember`/`recall`/etc. tool names; v2 internals live under `core/`, `memory/`, `retrieval/`.
+
+---
+
+# 23. Migration Strategy v1 → v2
+
+Migration must be non-destructive.
+
+## Phase M0 — Backup
+
+Before any schema migration:
+
+1. Verify DB exists.
+2. Create timestamped backup.
+3. Verify SQLite integrity.
+4. Record current schema version.
+
+## Phase M1 — Introduce migration metadata
+
+Create:
+
+```sql
+CREATE TABLE schema_meta (
+  key TEXT PRIMARY KEY,
+  value TEXT NOT NULL
+);
+```
+
+Store `schema_version`.
+
+## Phase M2 — Preserve old tables
+
+Do not immediately delete:
+
+```text
+preferences
+lessons
+interactions
+profile
+```
+
+## Phase M3 — Convert
+
+Map:
+
+```text
+preferences → memories(type=PREFERENCE)
+lessons → memories(type=LESSON)
+profile → profile projection/cache
+interactions → interactions unchanged
+```
+
+Preserve original IDs in metadata when IDs cannot be retained directly.
+
+## Phase M4 — Rebuild indexes
+
+Recreate FTS and semantic indexes from canonical v2 memory records.
+
+## Phase M5 — Validate
+
+Checks:
+
+- row counts
+- content hashes/sample comparisons
+- FTS availability
+- memory type mapping
+- profile availability
+- export availability
+
+Only after validation may v2 consider old tables deprecated.
+
+---
+
+# 24. Migration Files (as-built)
+
+Migrations are **TypeScript modules** in `src/db/migrations.ts`, not `.sql` files. Each entry is an idempotent `up(db)` (`CREATE TABLE IF NOT EXISTS`) tracked in `schema_meta`. This avoids `.sql` file-copy issues under `tsc` while keeping deterministic order. The logical schema in §6 is the contract; the TS implementation realizes it.
+
+Ordered migrations applied in v2.0.0:
+
+1. `schema_meta` table
+2. `memories` + indexes
+3. `entities` / `relations`
+4. `memory_links`
+5. v1 backfill (`preferences → PREFERENCE`, `lessons → LESSON`, sync FTS + embeddings; guarded by `v1_backfilled` flag, runs once)
+
+The spec allowed implementation differences ("Exact SQL can differ if implementation constraints require it"); the TS approach is the chosen realization.
+
+---
+
+# 25. Testing Strategy
+
+Tests must be layered.
+
+## Unit tests
+
+- classification
+- secret filtering
+- normalization
+- duplicate detection
+- conflict classification
+- confidence
+- decay
+- scoring
+- RRF
+- token budgeting
+- filename sanitization
+
+## Integration tests
+
+- SQLite migration
+- FTS sync
+- semantic retrieval
+- graph relations
+- lifecycle transitions
+- import/export
+
+## MCP E2E
+
+Run real JSON-RPC against the built server.
+
+## Plugin tests
+
+- prompt capture
+- tool event capture
+- dedupe
+- secret filter
+- graceful DB failure
+- profile/context injection
+
+---
+
+# 26. Retrieval Benchmark
+
+Create:
+
+```text
+benchmark/
+├── datasets/
+├── retrieval/
+├── conflict/
+├── temporal/
+├── lifecycle/
+└── performance/
+```
+
+Baseline dataset should contain:
+
+- at least 500 memories
+- at least 100 distractors
+- at least 100 duplicates
+- at least 100 contradictions/updates
+- at least 100 temporal changes
+
+Measure:
+
+```text
+Recall@1
+Recall@5
+Recall@10
+Precision@5
+MRR
+NDCG
+```
+
+Initial engineering targets:
+
+```text
+Recall@5 >= 0.90
+Precision@5 >= 0.85
+MRR >= 0.85
+```
+
+Targets are project acceptance goals, not claims about competitor performance.
+
+---
+
+# 27. Conflict Benchmark
+
+Target:
+
+```text
+>= 95% correct classification/resolution
+```
+
+Test cases must include:
+
+- exact duplicate
+- paraphrase duplicate
+- preference update
+- direct contradiction
+- temporary exception
+- two valid but different scoped memories
+- ambiguous conflict
+
+Ambiguous cases must prefer preservation over destructive guessing.
+
+---
+
+# 28. Temporal Benchmark
+
+Queries must test:
+
+```text
+current truth
+historical truth
+change detection
+supersession chain
+```
+
+No stale record may override an active current record in default current-context retrieval.
+
+---
+
+# 29. Performance Targets
+
+On a normal local development machine, initial targets are:
+
+| Operation | Target |
+|---|---:|
+| remember | <20 ms typical |
+| recall | <50 ms typical |
+| get_context | <100 ms typical |
+| get_profile | <20 ms typical |
+| search_history | <30 ms typical |
+
+These are engineering targets, not guaranteed SLAs.
+
+Benchmark both cold-cache and warm-cache behavior where practical.
+
+---
+
+# 30. Token Efficiency
+
+Every retrieval operation must have a bounded output.
+
+`get_context` must support an explicit token/character budget.
+
+Do not return all matching memories merely because they match.
+
+Target behavior:
+
+```text
+candidate pool: 50
+ ↓
+rank: 20
+ ↓
+filter: 12
+ ↓
+budget: 5–15 useful memories
+ ↓
+compact context
+```
+
+---
+
+# 31. Consolidation
+
+Consolidation creates higher-level semantic memories from clusters of related evidence.
+
+Example:
+
+```text
+User prefers TypeScript.
+User chooses TypeScript for projects.
+User corrected code examples from Python to TypeScript.
+```
+
+Can produce:
+
+```text
+User prefers TypeScript for software projects.
+```
+
+The derived memory must retain provenance:
+
+```text
+derived_from → source memories
+```
+
+Source evidence must not be deleted automatically.
+
+---
+
+# 32. Decision Memory
+
+`DECISION` should support optional rationale and alternatives in metadata.
+
+Example:
+
+```json
+{
+  "type": "DECISION",
+  "content": "Use SQLite for local persistence",
+  "metadata": {
+    "reason": [
+      "local-first",
+      "simple deployment",
+      "sufficient performance"
+    ],
+    "alternatives": ["PostgreSQL", "Neo4j"]
+  }
+}
+```
+
+This prevents agents from repeatedly reopening already-settled architecture decisions.
+
+---
+
+# 33. Scope Resolution
+
+Supported scope hierarchy:
+
+```text
+GLOBAL
+  ↓
+USER
+  ↓
+PROJECT
+  ↓
+SESSION
+```
+
+For current project queries, prefer:
+
+```text
+PROJECT > USER > GLOBAL > ARCHIVED
+```
+
+Session-specific temporary information should not silently become global memory.
+
+---
+
+# 34. Compatibility Requirements
+
+v2 must provide a migration/compatibility period where old workflows continue to work.
+
+Minimum compatibility:
+
+- existing `remember` usage
+- existing `recall` usage
+- existing `get_profile`
+- existing `save_lesson`
+- existing `search_history`
+- existing `forget`
+- existing `memory_stats`
+- existing `get_recent_interactions`
+- existing `export_memory`
+- existing OpenCode plugin DB path behavior
+
+Where behavior changes, document it explicitly in `MIGRATION_v2.md`.
+
+---
+
+# 35. Implementation Phases
+
+## Phase 0 — Freeze v1
+
+- tag v1.2.2
+- backup database
+- record baseline benchmarks
+- do not modify production behavior
+
+## Phase 1 — Core abstraction
+
+- repository layer
+- unified memory type
+- schema metadata
+- migration engine
+- v1 compatibility wrappers
+
+## Phase 2 — Lifecycle
+
+- status
+- confidence
+- importance
+- salience
+- access tracking
+- decay
+- archive
+- supersession
+
+## Phase 3 — Temporal
+
+- validity intervals
+- historical retrieval
+- change/supersession chains
+
+## Phase 4 — Conflict
+
+- normalization
+- duplicate detection
+- contradiction detection
+- update/supersession
+- merge
+
+## Phase 5 — Retrieval
+
+- FTS adapter
+- vector adapter
+- RRF
+- scoring
+- reranking
+
+## Phase 6 — Graph
+
+- entities
+- relations
+- memory links
+- bounded traversal
+- graph boost
+
+## Phase 7 — Context
+
+- `get_context`
+- token budgeting
+- context assembler
+- persistent/archival projection
+
+## Phase 8 — Consolidation
+
+- clustering
+- derived memories
+- provenance
+- optional AI-assisted summarization
+
+## Phase 9 — Benchmark/security
+
+- benchmark suite
+- migration tests
+- security tests
+- performance tests
+
+## Phase 10 — v2 release
+
+- v2 documentation
+- migration guide
+- changelog
+- package version 2.0.0
+
+---
+
+# 36. AI Coding Agent Rules
+
+This section is normative.
+
+## MUST
+
+- Read this document before modifying architecture.
+- Inspect current source before changing behavior.
+- Preserve v1 functionality unless explicitly superseded.
+- Add migrations instead of destructive schema replacement.
+- Add tests with every new subsystem.
+- Keep MCP stdio stdout clean; diagnostics belong on stderr.
+- Keep outputs bounded.
+- Preserve graceful failure.
+- Keep local/offline operation functional.
+- Treat stored memory as untrusted data.
+- Prefer deterministic logic over unnecessary LLM calls.
+
+## MUST NOT
+
+- Rewrite the entire project without migration.
+- Delete the v1 database schema before successful migration.
+- Introduce a mandatory cloud dependency.
+- Introduce a mandatory external embedding API.
+- add dozens of MCP tools for internal implementation details.
+- Let stale/superseded memories silently override current truth.
+- Destroy contradictory evidence merely because it is inconvenient.
+- Put secrets into test fixtures, logs, or examples.
+- Make plugin failure crash the host.
+- Change public behavior without tests and migration notes.
+
+## SHOULD
+
+- Keep modules small and independently testable.
+- Use interfaces/adapters for vector and graph implementations.
+- Prefer SQLite-native capabilities before adding dependencies.
+- Measure retrieval quality before tuning scoring constants.
+
+---
+
+# 37. Acceptance Criteria for v2.0.0
+
+The release is acceptable only when all are true:
+
+### Data
+
+- [ ] v1 DB can be backed up and migrated.
+- [ ] preferences map correctly to PREFERENCE memories.
+- [ ] lessons map correctly to LESSON memories.
+- [ ] interactions remain queryable.
+- [ ] profile remains available as a projection.
+
+### Memory
+
+- [ ] unified memory model works.
+- [ ] lifecycle states work.
+- [ ] temporal validity works.
+- [ ] supersession works.
+- [ ] duplicate detection works.
+- [ ] conflict handling preserves ambiguous evidence.
+
+### Retrieval
+
+- [ ] FTS retrieval works.
+- [ ] local semantic retrieval works.
+- [ ] RRF fusion works.
+- [ ] metadata/project scope works.
+- [ ] stale/superseded filtering works.
+- [ ] bounded output works.
+
+### Context
+
+- [ ] `get_context` works.
+- [ ] token/character budget is enforced.
+- [ ] current project context is prioritized.
+- [ ] critical constraints are prioritized.
+
+### Graph
+
+- [ ] entity/relation persistence works.
+- [ ] memory links work.
+- [ ] graph traversal is bounded.
+- [ ] graph failure does not break retrieval.
+
+### Security
+
+- [ ] secrets are filtered before storage.
+- [ ] exports are confined to the allowed directory.
+- [ ] memory cannot become executable instructions.
+- [ ] import validates schema/version.
+
+### Reliability
+
+- [ ] DB errors do not crash the MCP server.
+- [ ] plugin errors do not crash the host.
+- [ ] stdout remains protocol-safe.
+
+### Quality
+
+- [ ] retrieval benchmark passes project targets.
+- [ ] conflict benchmark meets >=95% target.
+- [ ] migration tests pass.
+- [ ] performance targets are measured and documented.
+
+---
+
+# 38. Recommended v2 Positioning
+
+Do not market v2 as "another Mem0".
+
+Position it as:
+
+> **A local-first, privacy-focused, temporal memory MCP for AI coding agents, with hybrid retrieval, conflict-aware memory, and token-efficient context assembly.**
+
+The differentiators are:
+
+1. Local-first.
+2. SQLite simplicity.
+3. No mandatory API/cloud.
+4. Thai/English friendliness.
+5. Cross-harness portability.
+6. Temporal + conflict-aware memory.
+7. Small MCP interface.
+8. Agent-oriented context assembly.
+
+---
+
+# 39. Final Architecture Contract
+
+The canonical v2 flow is:
+
+```text
+                    ┌──────────────────┐
+                    │    AI AGENT      │
+                    └────────┬─────────┘
+                             │ MCP
+                    ┌────────▼─────────┐
+                    │  MEMORY API      │
+                    └────────┬─────────┘
+                             │
+          ┌──────────────────┼──────────────────┐
+          │                  │                  │
+          ▼                  ▼                  ▼
+       CAPTURE          RETRIEVAL          LIFECYCLE
+          │                  │                  │
+          │        ┌─────────┼─────────┐        │
+          │        │         │         │        │
+          │       FTS      VECTOR     GRAPH     │
+          │        │         │         │        │
+          │        └─────────┼─────────┘        │
+          │                  │                  │
+          └──────────────────┼──────────────────┘
+                             ▼
+                    ┌────────────────┐
+                    │ MEMORY STORE   │
+                    │ SQLite + FTS5  │
+                    │ vectors + graph│
+                    │ temporal state │
+                    └───────┬────────┘
+                            │
+                    ┌───────▼────────┐
+                    │ CONTEXT ENGINE │
+                    │ rank/filter    │
+                    │ dedupe/compress│
+                    │ token budget   │
+                    └───────┬────────┘
+                            │
+                            ▼
+                         AI AGENT
+```
+
+**This document is the implementation source of truth for th-memory-mcp v2 unless a later version explicitly supersedes it.**
+
+---
+
+# 40. Implementation Status (as-built, v2.1.0)
+
+This section folds in the former `design.md` build log. All v2 engine phases are complete and tested. v2.1.0 adds the five previously-deferred tools.
+
+## What shipped (v2.0.0 + v2.1.0)
+- **16 MCP tools** (see §17). `save_lesson` retained as a compatibility wrapper. The four tools originally marked deferred (`update_memory`, `merge_memory`, `link_memory`, `import_memory`) plus `extract_memories` landed in v2.1.0.
+- **Non-destructive migration** from v1.2.2: v1 tables preserved; `memories`/`entities`/`relations`/`memory_links`/`schema_meta` added; one-time backfill of preferences + lessons.
+- **Hybrid retrieval**: FTS5 + local semantic (hashing-trick vectors) fused via RRF, then scored by confidence × importance × recency × scope.
+- **Temporal model**: validity intervals, point-in-time retrieval, supersession chains, change detection.
+- **Conflict/dedup**: normalize → exact → similar → classify (duplicate/update/contradiction/unrelated); ambiguous conflicts preserved, never silently destroyed.
+- **Graph**: entities/relations + bounded memory-link traversal (maxDepth 1–5); `link_memory` exposes it publicly.
+- **Context engine**: `get_context` with token budgeting, temporal filtering, optional graph expansion.
+- **Consolidation**: clustering + derived memories with `derived_from` provenance.
+- **Auto-extraction**: `extract_memories` scans recent interactions for memory-intent (deterministic, no LLM) and proposes/creates memories.
+- **Security**: secret filtering, parameterized SQL, FTS-injection quoting, memory treated as untrusted data, bounded output.
+
+## Phase checklist
+- [x] Phase 1 — Core abstraction (types, migrations, repository, index wiring)
+- [x] Phase 2 — Lifecycle engine (decay, source-weights, scorer, transitions)
+- [x] Phase 3 — Temporal model
+- [x] Phase 4 — Conflict & dedup (+ fixed v1 `embed.ts` DataView bug)
+- [x] Phase 5 — Hybrid retrieval (FTS/vector/fusion/scorer/engine)
+- [x] Phase 6 — Graph engine
+- [x] Phase 7 — Context engine (`get_context`)
+- [x] Phase 8 — Consolidation (`consolidate`)
+- [x] Phase 9 — Benchmark + security suites (13 test suites total)
+- [x] Phase 10 — Docs + v2.0.0 release (npm, GitHub Release, Official MCP Registry, Glama)
+- [x] v2.1.0 — `link_memory` / `merge_memory` / `update_memory` / `import_memory` / `extract_memories` (16 tools, 13 suites)
+
+## Test status
+All 13 test suites pass (capture, distill, lifecycle 17, temporal 7, conflict 14, retrieval 7, graph 7, context 7, consolidation 5, benchmark 2, security 5, tools_v21 21, smoke 16-tool).
+
+## Known gaps vs original spec (deferred, not regressions)
+- Retrieval/conflict benchmark quality targets (Recall@5 ≥ 0.90 etc., §26/§27) are **not yet measured** in-repo; only a performance benchmark (retrieve < 2000 ms over 300 memories) exists.
+- Perf targets in §29 are engineering goals, not yet benchmarked in CI.
+- `extract_memories` is a deterministic heuristic extractor (no LLM); AI-assisted extraction/summarization remains future work.
+- Auto entity extraction in consolidation is future work.
+
+## Release
+- v2.0.0: released — npm `th-memory-mcp@2.0.0` (latest), GitHub Release `v2.0.0`, Official MCP Registry `io.github.worakorn-prince/th-memory-mcp@2.0.0`, Glama listed.
+- v2.1.0: implemented and tested locally; **npm/GitHub/MCP Registry/Glama publish pending** because the publish token expired and requires the owner to re-authenticate (`npm login` / `mcp-publisher` GitHub OAuth).