Self-evolving AI memory system with three-tier architecture: session → short-term → long-term. Dual-path search (BM25 + vector), graph-enhanced retrieval for associative recall, and an evolution engine that continuously accumulates, distills, and corrects knowledge — delivered as an Agent Skill compatible with any AI assistant.
- Three-Tier Memory — Session (conversation buffer) → Short-term (round-based TTL, auto-decay) → Long-term (episodic / semantic / rules)
- Dual-Path Search — BM25 keyword + vector similarity, fused via Reciprocal Rank Fusion (RRF)
- Graph-Enhanced Retrieval — Explicit edges (provenance, semantic, causal, conflict) + implicit vector similarity for associative recall
- Evolution Engine — 8-step workflow promoting short-term observations into long-term knowledge with graph provenance tracking
- Context Bootstrap —
mem-x recallaggregates rules, tasks, recent memories, and graph edges for cold-start - Multi-Bucket Isolation — Each agent gets an isolated data directory under
~/.mem-x/<bucket>/ - Debug Dashboard — Browser-based UI for inspecting memory stats, timeline, health, and search
- Pluggable Embedding — OpenAI API, Ollama, or any OpenAI-compatible endpoint (e.g. LM Studio)
- Universal Agent Skill — A single
SKILL.mdgrants any AI assistant self-evolving memory - SQLite All-in-One — Data, FTS5 full-text index, and sqlite-vec vector index in a single file
bun install
bun run build
# Initialize database (creates ~/.mem-x/default/mem-x.db)
npx mem-x init
# Configure embedding (stored in ~/.mem-x/config.json)
npx mem-x config set embedding.apiKey sk-xxx
npx mem-x config set embedding.baseUrl http://localhost:1234/v1
npx mem-x config set embedding.model text-embedding-bge-m3# Database
mem-x init # Initialize database
# Context Bootstrap
mem-x recall [--limit N] # Aggregate rules + tasks + recent memories
# Session Memory (Tier 1 — conversation buffer)
mem-x session start # Start a new session
mem-x session add <id> --content "..." [--tags] # Add entry to session
mem-x session end <id> [--ttl <rounds>] # End session → commit to short-term
mem-x session show <id> # Show session details
mem-x session list # List recent sessions
# Memory CRUD (Tier 2 & 3)
mem-x memory add short_term --content "..." [--tags "a,b"] [--ttl <rounds>]
mem-x memory add episodic --event "..." [--context "..."] [--result "..."] [--tags "a,b"]
mem-x memory add semantic --topic "..." --content "..." [--tags "a,b"]
mem-x memory add rules --trigger "..." --constraint "..." [--reason "..."]
mem-x memory list <layer> [--since YYYY-MM-DD] [--limit N]
mem-x memory get <id> [--layer <layer>]
mem-x memory delete <id> [--layer <layer>]
mem-x memory purge # Remove expired short-term memories
# Search (dual-path: BM25 + vector, optional graph expansion)
mem-x search "<query>" [--layer short_term|episodic|semantic|rules] [--mode bm25|vector|hybrid] [--limit N]
mem-x search "<query>" --graph # Graph-enhanced: expand neighbors + boost connected
mem-x search "<query>" --graph --graph-depth 2 # 2-hop graph expansion
# Graph (memory relationships)
mem-x graph link <src> <tgt> --relation <type> --source-layer <L> --target-layer <L> [--weight N]
mem-x graph unlink <edge-id> # Remove an edge
mem-x graph neighbors <id> [--relation <type>] # List connected memories
mem-x graph list [--relation <type>] [--layer L] # List all edges
mem-x graph auto-link [--threshold N] # Auto-discover similar_to edges via vector KNN
# Tasks
mem-x task add --title "..." [--deadline "..."] [--priority low|medium|high|urgent]
mem-x task list [--status pending|in_progress|done|cancelled]
mem-x task update <id> --status <status>
# Debug Dashboard
mem-x debug [--port 3030] # Launch browser-based debug UI
# Config
mem-x config show
mem-x config set <key> <value>
# Multi-Bucket (agent isolation)
mem-x --bucket my-agent <command> # Use isolated data directory under ~/.mem-x/
MEM_X_BUCKET=my-agent mem-x <command> # Or via environment variable┌──────────────────────────────────────────────────────────┐
│ Agent Skill │
│ (skills/mem-x/SKILL.md) │
├──────────────────────────────────────────────────────────┤
│ CLI Layer │
│ init│session│memory│search│task│recall│debug│graph │
├──────────────────────────────────────────────────────────┤
│ Session (JSON) │ Short-term + Long-term (SQLite) │
│ ~/.mem-x/ │ Memory Store │ Dual-Path Search │
│ <bucket>/ │ CRUD + embed │ BM25 + Vector → RRF │
│ │ │ + Graph Expansion │
├──────────────────────────────────────────────────────────┤
│ Graph Layer (edges table) │ SQLite Database │
│ Provenance / Semantic / Causal│ FTS5 + sqlite-vec │
│ Conflict / Auto-similar │ │
├──────────────────────────────────────────────────────────┤
│ Embedding Providers │ │
│ OpenAI / Ollama / Custom │ │
└──────────────────────────────────────────────────────────┘
| Tier | Layer | TTL | Storage | Search |
|---|---|---|---|---|
| 1 | Session | Conversation | JSON files | N/A |
| 2 | Short-term | 7 rounds (configurable) | SQLite + FTS5 + vec0 | BM25 + Vector |
| 3 | Rules (highest priority) | Permanent | SQLite + FTS5 + vec0 | BM25 + Vector |
| 3 | Semantic | Permanent | SQLite + FTS5 + vec0 | BM25 + Vector |
| 3 | Episodic | Permanent | SQLite + FTS5 + vec0 | BM25 + Vector |
Session entries ──[session end]──▶ Short-term ──[evolution]──▶ Episodic
──▶ Semantic
──▶ Rules
──▶ (discard)
Graph Layer (edges) connects memories across all layers:
promoted_from │ derived_from │ related_to │ contradicts
supersedes │ caused_by │ leads_to │ similar_to
src/
├── cli.ts # Entry point + global --bucket option
├── cli/ # CLI commands
│ ├── session.ts # session start/add/end/show/list
│ ├── memory.ts # memory list/get/delete/purge
│ ├── memory-add.ts # memory add (extracted subcommand)
│ ├── memory-utils.ts # Shared CLI helpers
│ ├── search.ts # search (BM25/vector/hybrid/graph)
│ ├── graph.ts # graph link/unlink/neighbors/auto-link
│ ├── task.ts # task add/list/update
│ ├── recall.ts # recall (context bootstrap)
│ ├── config.ts # config show/set
│ ├── init.ts # init
│ └── debug.ts # debug (dashboard server)
├── db/
│ ├── connection.ts # SQLite singleton + sqlite-vec extension
│ ├── schema/ # Schema definitions (split for clarity)
│ │ ├── tables.ts # Data tables + metadata
│ │ ├── fts.ts # FTS5 virtual tables + triggers
│ │ ├── vec.ts # vec0 vector tables
│ │ └── index.ts # Schema orchestrator
│ ├── migrate.ts # Database initialization
│ └── rounds.ts # Round counter for short-term TTL
├── debug/
│ ├── handlers.ts # Stats, timeline, health, inspect
│ ├── search-debug.ts # Debug-mode search with score breakdown
│ └── server.ts # HTTP server + dashboard HTML
├── embedding/
│ ├── provider.ts # EmbeddingProvider interface
│ ├── openai.ts # OpenAI-compatible provider
│ ├── ollama.ts # Ollama provider
│ └── factory.ts # Provider factory
├── graph/
│ ├── types.ts # Edge/EdgeRelation interfaces
│ ├── edges.ts # Edge CRUD operations
│ ├── traverse.ts # Graph traversal (BFS, neighbors)
│ ├── auto-link.ts # Cross-layer vector KNN similarity
│ ├── enhance.ts # Graph-enhanced search (expand + boost)
│ └── index.ts # Barrel re-exports
├── memory/
│ ├── types.ts # TypeScript interfaces
│ ├── session.ts # Session memory (JSON file I/O)
│ ├── layers/ # Memory layer CRUD (one file per layer)
│ │ ├── episodic.ts
│ │ ├── semantic.ts
│ │ ├── rules.ts
│ │ ├── short-term.ts
│ │ └── tasks.ts
│ ├── helpers.ts # Tags, embedding, row hydration
│ ├── store.ts # Barrel re-exports
│ ├── search.ts # BM25 + vector + RRF fusion
│ └── recall.ts # Context bootstrap aggregation
└── utils/
├── config.ts # Configuration management (~/.mem-x/config.json)
├── bucket.ts # Multi-bucket path resolution (~/.mem-x/<bucket>/)
└── id.ts # UUID generation
The skills/mem-x/SKILL.md file is the complete AI skill definition. Point any AI assistant (Cursor, Claude Code, Windsurf, etc.) to this file and it will gain self-evolving memory capabilities — no code integration needed.
A Chinese version is available at skills/mem-x/SKILL_ZH.md.
- Runtime: Node.js >= 20
- Package Manager: Bun
- Language: TypeScript (ESM)
- Database: SQLite (better-sqlite3) + FTS5 + sqlite-vec
- Embedding: OpenAI SDK (compatible with LM Studio, Ollama, etc.)
- Testing: Vitest (186 tests, 90%+ coverage)
- Linting: ESLint + typescript-eslint
MIT