a-memory¶
Product name of the
mcp-ariel-memoryserver — your AI agents forget; a-memory makes them remember.
Universal Two-Layer Memory MCP Server for AI agents
What is it?¶
mcp-ariel-memory is a production-ready MCP server providing persistent, searchable memory for AI agents. It implements a two-layer architecture:
- Layer 1 (User) — facts about users: preferences, conversation history, emotional context
- Layer 2 (Agent) — agent identity: decisions, errors, personality evolution
Key Features¶
| Feature | Description |
|---|---|
| 5 universal primitives | think / dream / forget / evolve / project; full layer-tool surface behind ARIEL_EXPOSE=all |
| 4-layer memory | L1 ReflexBuffer → L2 Episodic → L3 Session → L4 Core |
| Typed memory | 13 categories with per-type retention, decay, and boost |
| RAG search | FTS5 + binary embeddings + hybrid scoring |
| Knowledge graphs | Epistemic (facts/decisions) + Temporal (timeline) |
| Wiki system | .md files as source of truth, 14 content types |
| Saga pattern | Multi-step ops with retry, idempotency, compensation |
| Envelope encryption | libsodium secretbox, keychain-first key resolution |
| Platform-aware async | aiosqlite on Linux/macOS, asyncio.to_thread on Windows |
| SHA-256 dedup | Prevents duplicate observations within 5-minute window |
| Circuit breaker | Prevents cascading LLM/embedding failures |
| Token budget | Limits context injection to 2000 tokens |
| Privacy filter | Strips API keys, secrets, and private tags |
Quick Start¶
Claude Desktop¶
Documentation¶
| Section | Description |
|---|---|
| Architecture | Layered model, L1-L4, consolidation, 23 DB tables |
| MCP Tools | Tool reference (layer ops; agents normally see only the 5 primitives) |
| RAG & Search | Unified search, BM25 conflict similarity, type-aware boost |
| Hooks | 19 lifecycle hooks (12 user + 13 agent slots), importance gating |
| Operations | Transports, health, auth, configuration |
| API Reference | Auto-generated from docstrings |
Status¶
- Version: 1.5.0
- Tests: 400+ passed (including 25 property-based Hypothesis tests)
- DB tables: 23
- Python: 3.10–3.13
- Platform: Windows, Linux, macOS, Docker