AI Memory MCP Server

Agent-agnostic persistent memory as an MCP Server — local-first: memories travel with your project in .aamm/, shared across Claude Code / Qoder / Cursor.

中文 | 日本語

An agent-agnostic persistent memory layer exposed as an MCP Server. Any MCP client — Claude Code, Qoder, Cursor — can reuse it. Memories live in each project's .aamm/ directory and travel with the project; different agents working the same project share one memory store, with a source_agent stamp distinguishing writers.

Architecture

• SQLite — structured source of truth (CRUD + FTS5 keyword search)
• Chroma (embedded) — vector retrieval, persisted to .aamm/chroma/
• Embedding — any OpenAI-compatible service (Volcengine / SiliconFlow / OpenAI / others); defaults to Volcengine doubao-embedding-vision
• Markdown mirror — each memory is also written to .aamm/memories/<category>/<id>.md, human-readable and editable

The three layers are joined by id.

Memory categories

Install

From PyPI:

pip install ai-agent-memory-mcp

From source:

cd ai_agent_memory_mcp
pip install .          # or: pip install -e .   (editable, for development)

Requires Python 3.11+.

Configure embedding (any OpenAI-compatible service)

The embedding layer is a generic OpenAI-compatible client — Volcengine / SiliconFlow / OpenAI / any compatible service works. On first run a default config is generated at .aamm/config.yml; edit as needed.

Fields (embedding section of .aamm/config.yml)

Put the key in the project root .env, then edit the embedding section of config.yml.

Examples

Volcengine doubao-embedding-vision (default; Agent/Coding Plan keys must use the Plan endpoint /api/plan/v3 — standard /api/v3 returns 401)

embedding:
  provider: volcengine
  model: doubao-embedding-vision
  base_url: https://ark.cn-beijing.volces.com/api/plan/v3
  api_key_env: VOLCENGINE_API_KEY
  dim: 2048

.env: VOLCENGINE_API_KEY=...

SiliconFlow bge-large-zh (Chinese-text optimized)

embedding:
  provider: siliconflow
  model: BAAI/bge-large-zh-v1.5
  base_url: https://api.siliconflow.cn/v1
  api_key_env: SILICONFLOW_API_KEY
  dim: 1024

.env: SILICONFLOW_API_KEY=...

OpenAI

embedding:
  provider: openai
  model: text-embedding-3-small
  base_url: https://api.openai.com/v1
  api_key_env: OPENAI_API_KEY
  dim: 1536

.env: OPENAI_API_KEY=...

Any other OpenAI-compatible service: just fill in base_url / model / api_key_env / dim.

After switching embedding model, old vectors may mismatch in dimension; clear .aamm/chroma/ and re-remember, or run python tests/rebuild_vectors.py.

Retrieval

recall uses three-way fused retrieval to maximize hit rate:

• Vector (weight 0.6): Chroma cosine; embeddings are computed from title + tags + content, so title/tag signal enters the vector
• Keyword (weight 0.25): SQLite FTS5 trigram
• Title/tag match (weight 0.15): +0.15 if the query appears in the title, +0.075 if in a tag

Candidates are expanded to top_k*3, then fused down to top_k. If the query contains FTS5 special characters (., *, ", -, ...), the keyword branch falls back to LIKE substring matching instead of erroring.

Work journal

Besides searchable memories, aamm keeps a human-readable work journal. After completing a user request, the agent calls journal_entry() to log what was asked / what it did / any open question. Journals are for people reading a timeline; recall does not search them. Use search_journal() only as a fallback to recover "what happened in a past interaction".

Journals are written to .aamm/logs/:

• journal.db — single SQLite store (the search source, spans all dates)
• YYYY-MM-DD.md — one Markdown file per day, append-only timeline

.aamm/logs/
├── journal.db        # search source (all dates)
├── 2026-07-14.md     # per-day timeline
└── 2026-07-15.md

MCP tools

Memory (8):

• remember(title, content, category, tags?, scope?) — store (three-way sync, auto-embed)
• recall(query, category?, top_k=5) — fused retrieval (vector + keyword + title match)
• get_memory(id) — get one
• search_memories(category?, tag?, agent?) — structured filter
• update_memory(id, ...) — update (re-embed + refresh md)
• forget(id) — delete (three-way sync)
• list_memories(category?) — list
• who_am_i() — current agent + project context

Journal (3):

• journal_entry(question, answer_summary, key_points?, open_question?, session_id?) — log a timeline entry
• search_journal(query, date_from?, date_to?, agent?, limit=10) — fallback search over journals
• setup_profile(user_name) — set the user name (shown in journals)

Management CLI

python -m ai_agent_memory_mcp.cli init                  # initialize .aamm in the current project
python -m ai_agent_memory_mcp.cli status                # store overview (categories / vectors / md / journal)
python -m ai_agent_memory_mcp.cli export [--dir DIR]    # export all memories to Markdown
python -m ai_agent_memory_mcp.cli sync                  # rebuild SQLite + Chroma from Markdown
python -m ai_agent_memory_mcp.cli check                 # consistency check (db / md / chroma)
python -m ai_agent_memory_mcp.cli journal [--limit N]   # show recent journal entries

Wire into Claude Code (user scope; shared code, per-project data)

From PyPI (no PYTHONPATH needed):

claude mcp add aamm -s user -- python -m ai_agent_memory_mcp --agent claude-code --project-from-cwd

From a source clone, add -e PYTHONPATH=<clone dir>\ai_agent_memory_mcp:

claude mcp add aamm -s user -e PYTHONPATH=<clone dir>\ai_agent_memory_mcp -- python -m ai_agent_memory_mcp --agent claude-code --project-from-cwd

Qoder / Cursor are the same — just change --agent.

Data layout

.aamm/
├── memory.db                    # SQLite: structured memories + FTS5
├── chroma/                      # Chroma vector store
├── memories/<category>/<id>.md  # Markdown mirror (editable)
├── logs/
│   ├── journal.db               # work journal (search source)
│   └── YYYY-MM-DD.md            # per-day journal timeline
├── config.yml                   # embedding config
└── profile.json                 # user name

License

MIT