AGENTS.md - mnemo

MCP Server cho AI memory. Python 3.13, uv, hatchling, src layout. Post-de-host (2026-09): server la MOT HTTP Streamable endpoint tren hull-core (auth no-auth|token|multi + per-namespace SQLite), local SQLite (WAL) la storage authority duy nhat. Khong con stdio transport, Cloudflare deploy, passport sync, setup form, hay multi-provider key dispatch. Hybrid search: FTS5 + sqlite-vec. 13 tools: 11 specialized memory tools (add_memory, search_memory, list_memories, update_memory, delete_memory, export_memories, import_memories, memory_stats, restore_memory, archived_memories, consolidate_memories) + legacy memory dispatcher (DEPRECATED -- use the granular tools instead) + config. 18 memory actions. Provider calls di qua hull per-task [models.*] cells (base_url + api_key + model, OpenAI-spec HTTP; OpenRouter pre-wired default): [models.embed] (cloud embed; local Qwen3 ONNX fallback), [models.rerank] (cloud rerank; local cross-encoder fallback), [models.chat] (compression, entity extraction, consolidation), [models.jev_score] (importance scoring). Cell khong co key => feature do skip gracefully, KHONG fallback sang provider khac.

Commands

# Setup
uv sync --group dev

# Lint & Type check
uv run ruff check .
uv run ruff format --check .
uv run ty check

# Fix
uv run ruff check --fix .
uv run ruff format .

# Test (integration excluded by default)
uv run pytest
uv run pytest tests/test_db.py -v                          # single file
uv run pytest tests/test_db.py::TestSearch::test_basic -v  # single test

# Build & Run
uv build
uv run n24q02m-mnemo                    # HTTP server (bind tu ~/.mnemo/config.toml)
uv run n24q02m-mnemo config-init        # ghi template config
uv run n24q02m-mnemo warmup             # pre-download local ONNX embed model

# Mise shortcuts
mise run setup     # full dev setup
mise run lint      # ruff check + format check + ty check
mise run test      # pytest
mise run fix       # ruff fix + format

Pytest

• asyncio_mode = "auto" -- khong can @pytest.mark.asyncio
• Timeout: 30s/test
• Integration marker: @pytest.mark.integration (can network/services)
• Default: -m 'not integration and not live and not full'
• Snapshot testing: syrupy

Cau truc thu muc

src/mnemo/
  __main__.py      # python -m mnemo entrypoint
  cli.py           # n24q02m-mnemo entry (alias mnemo-mcp): bare = HTTP server; subcommands
                   # (config-init, warmup, token-hash, token-verify)
  server.py        # FastMCP tools/resources/prompts + hull HTTP app
  runtime.py       # bridge sang hull-core: settings, model cells, auth,
                   # per-namespace DB paths
  llm.py           # completion dispatch qua [models.chat] cell
  capture.py       # typed capture pipeline (dedup + compression hook)
  compression.py   # LLM compression qua chat cell ("chat-cell" marker)
  db.py            # SQLite: CRUD, FTS5, vector search (sqlite-vec)
  embedder.py      # [models.embed] cell + fastretrieval local fallback
  reranker.py      # local Qwen3 cross-encoder + [models.rerank] cell chain
  graph.py         # entity/relation extraction qua chat cell
  temporal/        # bitemporal KG: extract, resolve, store, queries
  alembic/         # schema migrations (mem_001..mem_003)
  setup_tool.py    # warmup logic
  providers.py     # bounded paid reflect provider (cap USD)
  pilot_tools.py   # CLI-surface handlers over mnemo_core
  docs/            # tool documentation markdown (memory.md, config.md)
src/mnemo_core/    # domain core for the `mnemo` CLI surface:
                   # operations, standing pages, ports, results, defense
src/mnemo_cli/     # `mnemo` / `mnemo-pilot` argparse CLI (no server)
tests/             # 1:1 mapping voi source modules

Env vars

• MNEMO_DB_PATH / DB_PATH -- default ~/.mnemo/memories.db
• MNEMO_HOST / MNEMO_PORT -- HTTP bind overrides (config.toml [server])
• MNEMO_AUTH_TOKEN -- token source cho n24q02m-mnemo token-hash
• HULL_<TASK>_API_KEY -- host-only provider key override per cell: HULL_EMBED_API_KEY / HULL_RERANK_API_KEY / HULL_CHAT_API_KEY / HULL_JEV_SCORE_API_KEY (env wins over config.toml api_key)
• EMBEDDING_DIMS -- storage width (0 = runtime default 1024)
• REINDEX_ON_MODEL_CHANGE -- clear stale vector state on model swap
• DISABLE_LOCAL_EMBED / DISABLE_LOCAL_RERANK -- kill local ONNX fallback
• LOCAL_EMBEDDING_MODEL / LOCAL_RERANK_MODEL -- local model override
• RERANK_ENABLED, RERANK_TOP_N (10)
• ARCHIVE_ENABLED, ARCHIVE_AFTER_DAYS (90), ARCHIVE_IMPORTANCE_THRESHOLD (0.3), ARCHIVE_TRIGGER_EVERY (100)
• DEDUP_THRESHOLD (0.9), DEDUP_WARN_THRESHOLD (0.7)
• RECENCY_HALF_LIFE_DAYS (7)
• COMPRESSION_ENABLED (true)
• KG_AUTO_ENABLED (false), TEMPORAL_ENTITY_RESOLUTION_THRESHOLD (0.85), TEMPORAL_SUPERSESSION_THRESHOLD (0.85), TEMPORAL_SUPERSESSION_ENABLED (true)
• MNEMO_REFLECT_MODEL, MNEMO_REFLECT_CAP_USD -- paid reflect bounds
• FASTRETRIEVAL_CACHE_PATH, LOG_LEVEL (INFO)
• Removed 2026-09 (gone from code, do not reintroduce): EMBEDDING_MODELS, RERANK_MODELS, LLM_MODELS, *_API_BASE endpoints, per-provider *_API_KEY (JINA/GEMINI/OPENAI/COHERE/XAI/ANTHROPIC/VERTEX_EXPRESS), COMPRESSION_PROVIDER, COMPRESSION_MODEL, SYNC_*, GOOGLE_DRIVE_CLIENT_ID, MEMORY_DB_BACKEND, MCP_STORAGE_BACKEND, MCP_TRANSPORT, PUBLIC_URL, MCP_DCR_SERVER_SECRET, MNEMO_ENTERPRISE*, MNEMO_AUDIT_*. See docs/ARCHITECTURE.md history.

Client config

Server chi noi Streamable HTTP -- khong co stdio spawn. Chay instance roi point client vao endpoint:

uvx n24q02m-mnemo   # http://127.0.0.1:8000/mcp (no-auth default)

{
  "mcpServers": {
    "mnemo": { "type": "http", "url": "http://127.0.0.1:8000/mcp" }
  }
}

Provider cells song trong ~/.mnemo/config.toml (n24q02m-mnemo config-init): [models.<task>] = base_url + api_key + model, OpenRouter pre-wired. Key co the dat trong file hoac qua HULL_<TASK>_API_KEY; key do chi la seed cho key pool cua hull (~/.mnemo/credentials.db, rotate tren 401/403/429/5xx, named account qua account = "<id>").

Embedding architecture

1. Local -- Qwen3-Embedding ONNX via fastretrieval, zero config, default khi [models.embed] cell khong co key.
2. Cloud ([models.embed] cell) -- OpenAI-spec /embeddings call.

Local SQLite dung storage width mac dinh 1024 (native cua default embed cell). Never mix vectors from different model identities -- REINDEX_ON_MODEL_CHANGE=true clears stale vector state truoc next pass.

Storage authority

• SQLite (~/.mnemo/memories.db, WAL) la authority duy nhat. auth = "multi" tach per-namespace store ~/.mnemo/subs/<ns>/memories.db.
• Backup / cross-machine migration = rclone BEN NGOAI server; khong co embedded sync. Cloudflare D1/Vectorize va GDrive/S3 passport sync da removed 2026-09 (history trong git; ARCHITECTURE.md chi giu phan compression).

CD Pipeline

PSR v10 (workflow_dispatch) -> PyPI + GitHub Release; eligible stable releases -> MCP Registry + marketplace.

Luu y

• Tools tra ve _json({"error": "..."}), khong raise exception.
• match action: pattern cho routing trong memory dispatcher.
• asyncio.to_thread() cho blocking I/O (SQLite, embedding).
• Local embedding: first run download ~570MB model, cached.
• Dependencies: fastretrieval>=1.11.1, sqlite-vec, n24q02m-hull (PyPI, >=0.7.0,<0.8). Khong con litellm / mcp-core / native provider SDKs -- moi cloud call qua hull-core OpenAICompatClient tren per-task cells, xoay vong credential qua key pool cua hull (mnemo/keys.py, store ~/.mnemo/credentials.db).
• Pre-commit: gitleaks, ruff lint + format, ty check, pytest.
• Secrets: skret SSM namespace /mnemo/prod (region ap-southeast-1)

E2E

tests/live_http.py spawn that su HTTP server tren tmp HOME (config-init + env isolation); tests/test_live_protocol*.py, test_http_direct.py, test_live_mcp.py, test_full_live.py exercise protocol surface. Network-marked suites (integration/live/full/e2e) deselect by default. Khong con setup-form E2E hay mcp-core driver matrix cho repo nay.