mem-mesh 1.26.1

Centralized memory system for AI development tools

mem-mesh

Persistent memory for AI coding tools — resumable session state, the decision context that never reaches git, and injection that is measured rather than assumed. Hybrid vector + FTS5 search over a single SQLite file, zero external services.

한국어 · Quick Start · MCP Setup · MCP Tools · Session & Pins · Architecture · Docker · Contributing

---

Why mem-mesh?

mem-mesh does not claim that searching your past sessions makes a model write better code — it treats that as a hypothesis it instruments and measures (Measured, not assumed). It earns its place on three things git, pull requests, and well-kept docs do not capture:

• Session-to-session work state — pin_add / pin_complete track the unit of work; session_resume restores where the last session stopped. "Where was I" answered directly.
• Knowledge that never reaches git — the why of a decision, the approach that failed, the constraint learned during an incident. First-class categories with typed relations, so a superseded decision links to the one that replaced it.
• Observability and retrospective — weekly_review, the dashboard, and team relay show what your agents recorded, retrieved, and let go stale.

Differentiator	What it means
Pin lifecycle	Lightweight kanban inside every session: pin_add → pin_complete → pin_promote. Restore in-progress work with session_resume.
Injection instrumentation	Every auto-surfaced memory is tracked in injected_memories; a deterministic Stop-time heuristic judges whether it was used; weekly_review reports the hit rate. Utility is measured, not assumed.
Git-anchored staleness	Code memories carry commit + file anchors; the client verifies freshness and reports it, so stale context is dropped from injection instead of misleading the agent.
Human-gated doc promotion	doc_proposal drafts a promotion to version-controlled docs; a person approves; the client applies it. Memory is the staging area, git is the durable layer.
Hybrid search	sqlite-vec vector embeddings + FTS5 full-text fused with Reciprocal Rank Fusion (RRF). Korean n-gram optimized out of the box.
NLI conflict detection	2-stage pipeline: vector similarity pre-filter → mDeBERTa NLI model catches contradictory memories before they're stored.
4-Tier Smart Expand	session_resume(expand="smart") uses an importance × status matrix to load only what matters — ~60% token savings.
Zero external services	Single SQLite file. pip install mem-mesh and you're running. No Postgres, no Redis, no cloud.
Dual MCP transport	stdio (Cursor, Claude Desktop, Kiro) + Streamable HTTP/SSE (MCP spec 2025-03-26).
25+ client auto-detection	Identifies the calling IDE/AI platform from MCP handshake or User-Agent.
Batch operations	Pack multiple memory ops into one round-trip: 30–50% token savings.

Measured, not assumed

The "past-session search boosts coding" claim is instrumented rather than asserted. Every injected memory is recorded in injected_memories (one row per turn); a Stop-time heuristic with no LLM judges whether it was later referenced; weekly_review surfaces the injection hit rate. Offline, scripts/replay_injection_eval.py replays real captured prompts through the legacy and current injection formats and scores both with deterministic metrics plus an optional blind LLM judge. The premise is honest: if the current format shows no advantage, shrinking injection is a valid outcome. For pure code work in a repository whose commits, PRs, and docs are already well kept, the marginal value of retrieving past sessions is unproven — mem-mesh ships the tools to measure it, and the three capabilities above stand regardless of how that measurement resolves.

---

Features

• Memory CRUD — add, search, context, update, delete
• Hybrid search — sentence-transformers vectors + FTS5 RRF fusion, Korean n-gram support
• Session & pins — short-lived work tracking with importance-based promotion to permanent memory
• Injection instrumentation — injected_memories tracking + Stop-time usage heuristic + weekly_review injection stats; offline replay harness to validate injection value
• Git-anchored lifespan — commit/file anchors with client-side staleness verification; stale memories excluded from injection
• Human-gated promotion — doc_proposal promotes memory toward version-controlled docs (LLM drafts, human approves, client applies)
• Auto-redaction — deterministic secret/PII masking on auto-captured content before it reaches long-term memory
• Memory relations — link, unlink, get_links across 7 relation types
• Conflict detection — mDeBERTa NLI prevents storing contradictory facts
• Batch operations — 30–50% fewer tokens per multi-op workflow
• Web dashboard — FastAPI REST API + real-time UI at localhost:8000

---

Quick Start

Recommended: uvx (zero Python management)

One tool to install — uv — and mem-mesh handles the rest. No virtualenv, no pyenv tweaks, no sqlite-vec compile errors. Your MCP client spawns a cached, isolated mem-mesh on-demand.

# 1. Install uv (one-time, ~15 seconds)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Run the onboarding wizard — writes MCP config for detected tools,
#    offers to install hooks, warms the uv cache.
uvx mem-mesh

That's it. Restart Cursor / Claude Desktop / Kiro and mem-mesh MCP tools are live.

uvx mem-mesh (bare) runs onboarding — no --from "mem-mesh[server]" needed, since the lightweight base package is all the wizard requires (it still writes config that runs the server via the [server] extra). uvx mem-mesh install is the explicit equivalent.

For agents / CI: uvx mem-mesh --json runs onboarding non-interactively and prints a single JSON result (per-step status + next_actions). Non-TTY invocations (pipes, agents) auto-run non-interactively even without --json. Onboarding writes the config single-source-of-truth — ~/.mem-mesh/api_url and ~/.mem-mesh/hook_token — which every tool's hooks read at runtime (reachable from GUI- and terminal-launched tools alike). The MEM_MESH_API_URL / MEM_MESH_HOOK_TOKEN env vars seed those files (and act as a per-session override; mem-mesh doctor flags any that shadow the files). The token must also stay exported for HTTP hooks / authenticated MCP, which read the shell env and have no file fallback. Example: MEM_MESH_API_URL=https://memory.example.com MEM_MESH_HOOK_TOKEN=… uvx mem-mesh --json.

Install or repair hooks directly through the same uvx entrypoint:

uvx mem-mesh hooks install --target codex
uvx mem-mesh hooks status

Want the web dashboard too? uvx --from "mem-mesh[server]" mem-mesh serve — open http://localhost:8000.

Alternative: pip install

If you prefer managing Python environments yourself:

pip install "mem-mesh[server]"
mem-mesh                    # onboarding wizard (or: mem-mesh install)
mem-mesh serve             # web server + SSE MCP at localhost:8000

Prerequisites (only if NOT using uvx)

mem-mesh loads the sqlite-vec extension at runtime, so Python's sqlite3 module must support loadable extensions.

• uvx users — uv's managed Python builds already have extension loading enabled. Nothing to do.
• Linux — pysqlite3-binary wheel installs automatically as a fallback.
• macOS — system Python and Homebrew Python both work. Only pyenv's default build is broken.
• Windows — system Python works; install pysqlite3-binary manually if needed.

macOS + pyenv users who hit Migration failed: no such module: vec0:

# Option A: rebuild Python against Homebrew sqlite3
brew install sqlite3
SQLITE_PREFIX="$(brew --prefix sqlite3)"
PYTHON_CONFIGURE_OPTS="--enable-loadable-sqlite-extensions" \
LDFLAGS="-L${SQLITE_PREFIX}/lib" \
CPPFLAGS="-I${SQLITE_PREFIX}/include" \
CFLAGS="-I${SQLITE_PREFIX}/include" \
  pyenv install 3.13 --force
pyenv rehash

# Option B (simplest): just use uvx — it bypasses system Python entirely

Linux distro Python, Docker images, and conda Python ship with extension loading enabled — no extra steps needed.

---

MCP Setup

mem-mesh install writes these entries for you automatically. The snippets below are what gets written, for reference.

uvx (recommended)

Zero Python-env management. The MCP client spawns a cached mem-mesh process per call; the first run downloads it, subsequent runs are instant.

{
  "mcpServers": {
    "mem-mesh": {
      "command": "uvx",
      "args": ["--from", "mem-mesh[server]", "mem-mesh-mcp-stdio"],
      "env": { "MEM_MESH_CLIENT": "cursor" }
    }
  }
}

Stdio (local Python)

Use your own Python install. Good if you need -e . dev installs.

{
  "mcpServers": {
    "mem-mesh": {
      "command": "python",
      "args": ["-m", "app.mcp_stdio"],
      "cwd": "/absolute/path/to/mem-mesh",
      "env": { "MCP_LOG_LEVEL": "INFO" }
    }
  }
}

HTTP (streamable, shared running server)

For web clients or when multiple tools share one process. Requires mem-mesh serve running. Use "type": "http" — type: "sse" is legacy and hangs after a server restart.

{
  "mcpServers": {
    "mem-mesh": {
      "url": "http://localhost:8000/mcp/sse",
      "type": "http"
    }
  }
}

Config file locations by tool:

Tool	Config file
Cursor	.cursor/mcp.json
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json
Kiro	~/.kiro/settings/mcp.json

Mode comparison

	uvx	Stdio	HTTP
Prereq	uv only	Python env with mem-mesh[server]	Running mem-mesh serve
First call	~15s (cache warm)	Instant	Instant
Server to manage	None	None	Yes
Dashboard	Optional (uvx … serve)	Optional	Included
Hooks support	Requires separate server	Yes (local mode)	Yes (api mode)

---

MCP Tools (15)

Tool	Description	Key parameters
add	Store a memory	content, project_id, category, tags
search	Hybrid vector + FTS5 search	query, project_id, category, limit, recency_weight, response_format
context	Retrieve memories surrounding a given memory	memory_id, depth, project_id
update	Edit a memory	memory_id, content, category, tags
delete	Remove a memory	memory_id
stats	Usage statistics	project_id, start_date, end_date
link	Create a typed relation between memories	source_id, target_id, relation_type
unlink	Remove a relation	source_id, target_id
get_links	Query relations	memory_id, relation_type, direction
pin_add	Add a short-lived work-tracking pin	content, project_id, importance, tags
pin_complete	Mark a pin done; optionally promote to permanent memory	pin_id, promote, category
pin_promote	Promote an already-completed pin to permanent memory	pin_id, category
session_resume	Restore context from the previous session	project_id, expand, limit
session_end	Close a session with a summary	project_id, summary, auto_complete_pins
batch_operations	Execute multiple ops in one call	operations (array of add/search/pin_add/pin_complete)

search response formats: minimal | compact | standard | full

---

Search

mem-mesh runs two retrieval engines in parallel and merges results with Reciprocal Rank Fusion:

• Vector — dragonkue/snowflake-arctic-embed-l-v2.0-ko (1024-dim, Korean retrieval SOTA, MTEB-ko #1) by default; KURE, E5 and MiniLM models supported
• FTS5 — SQLite full-text search with n-gram tokenization for CJK languages
• RRF fusion — balances semantic similarity and keyword precision
• Quality filters — noise removal, intent analysis, vector pre-filter overfetch to improve recall

---

Session & Pins

Session lifecycle

session_resume(project_id, expand="smart")  →  work  →  session_end(project_id, summary)

• session_resume restores incomplete pins and context from the previous session. Stale pins are auto-closed. expand="smart" applies an importance × status matrix that cuts token usage by ~60%.
• session_end records a summary and closes the session. If the session terminates abnormally, the next session_resume automatically recovers open pins.

Pin lifecycle

Pins are the unit of work inside a session. Track code changes, implementations, and configuration work as pins — not in permanent memory.

pin_add(content, project_id)  →  do the work  →  pin_complete(pin_id, promote=True)
                                                   # promote=True completes and promotes in one call

Status flow: open (planned, not started) → in_progress (active; default on pin_add) → completed

Multi-step work can pre-register later steps as open pins, then activate them one at a time.

Auto-stale cleanup (triggered on session_resume):

• in_progress pins older than 7 days → auto-completed
• open pins older than 30 days → auto-completed

When to pin: only when files change. Questions, explanations, and read-only lookups do not need pins. Multi-step tasks get one pin per step.

Importance levels:

Level	Use for
5	Architecture decisions, core design changes
3–4	Feature implementations, significant fixes
1–2	Minor edits, typo fixes
omit	Auto-inferred from content

Promote: pin_complete(pin_id, promote=True) completes and promotes to permanent memory in one call. To promote after the fact: pin_promote(pin_id).

Client detection: In HTTP mode, the calling client is identified from the MCP initialize handshake or User-Agent header (25+ IDE/AI platforms supported). In stdio mode, set MEM_MESH_CLIENT in the environment.

AI agent checklist

1. Session start   → session_resume(project_id, expand="smart")
2. Past context    → search() before coding if referencing previous decisions
3. Track work      → pin_add → pin_complete (promote=True to merge into memory)
4. Permanent store → decision / bug / incident / idea / code_snippet only
5. Session end     → session_end(project_id, summary, auto_complete_pins=True)
6. Never store     → API keys / tokens / passwords / PII

Principle: Hooks are read-only signals. All pin creation, completion, and promotion decisions are made by the LLM with full context.

---

Memory Relations

Seven relation types: related | parent | child | supersedes | references | depends_on | similar

get_links direction: outgoing | incoming | both

---

Configuration

Variable	Description	Default
MEM_MESH_DATABASE_PATH	SQLite database path	XDG per-user path (see app/core/config.py _default_db_path)
MEM_MESH_EMBEDDING_MODEL	Embedding model name	dragonkue/snowflake-arctic-embed-l-v2.0-ko
MEM_MESH_EMBEDDING_DIM	Vector dimensions	1024
MEM_MESH_SERVER_PORT	Web server port	8000
MEM_MESH_SEARCH_THRESHOLD	Minimum similarity score	0.5
MEM_MESH_USE_UNIFIED_SEARCH	Enable hybrid search	true
MEM_MESH_ENABLE_KOREAN_OPTIMIZATION	Korean n-gram FTS	true
MEM_MESH_LOG_LEVEL	Server log level	INFO
MEM_MESH_LOG_FILE	Log output file	(none)

See .env.example for the full list.

Pointing hooks at a remote mem-mesh server

Bash hooks (Stop, SessionStart, SubagentStop, …) resolve the API URL in this order:

1. MEM_MESH_API_URL environment variable
2. API_URL environment variable
3. ~/.mem-mesh/api_url — single-line file with the server URL
4. The URL baked into the hook at install time
5. http://localhost:8000

Use the config file when you want one installed hook bundle to talk to a remote server (e.g. https://mem.example.com) without editing settings.json and without relying on env-var inheritance — Claude Code does not export settings.json.env retroactively to already-running sessions, so an env var added mid-session won't reach hooks until you restart.

mkdir -p ~/.mem-mesh
echo 'https://mem.example.com' > ~/.mem-mesh/api_url
mem-mesh doctor   # confirms the resolved URL and its source

Per machine — the file is not synced. Delete it (or set MEM_MESH_API_URL) to fall back to the baked default.

---

Web Dashboard

• Dashboard: http://localhost:8000
• API docs (Swagger): http://localhost:8000/docs
• Health check: http://localhost:8000/health

---

Architecture

flowchart LR
    subgraph Clients
        Cursor[Cursor]
        Claude[Claude Desktop]
        Kiro[Kiro]
        Web[Web Client]
    end

    subgraph Transport
        Stdio[Stdio MCP]
        SSE[SSE / Streamable HTTP]
    end

    subgraph Core
        MCP[mcp_common]
        Storage[Storage Service]
    end

    subgraph Data
        SQLite[(SQLite + sqlite-vec + FTS5)]
    end

    Cursor --> Stdio
    Claude --> Stdio
    Kiro --> Stdio
    Web --> SSE
    Stdio --> MCP
    SSE --> MCP
    MCP --> Storage
    Storage --> SQLite

Directory structure

mem-mesh/
├── app/
│   ├── core/              # DB, embeddings, services, schemas
│   ├── mcp_common/        # Shared MCP tools, dispatcher, batch
│   ├── mcp_stdio/         # FastMCP stdio server
│   ├── mcp_stdio_pure/    # Pure MCP stdio server
│   └── web/               # FastAPI (dashboard, SSE MCP, OAuth, WebSocket)
├── static/                # Frontend (Vanilla JS, Web Components)
├── tests/                 # pytest
├── scripts/               # Migration and benchmark scripts
├── docs/rules/            # AI agent rule modules
├── data/                  # memories.db
└── logs/

---

Docker

# Build and start
make quickstart
# or step by step:
make docker-build && make docker-up

# Open http://localhost:8000

---

First-run setup (dashboard auth)

When the server starts with no dashboard auth configured, anyone who can reach the port can read/write/delete all memories. To let you close this from the browser (no shell-only config), mem-mesh mints a one-time setup token on first boot and prints it to the server console:

============================================================
  FIRST-RUN SETUP  —  dashboard auth is NOT configured
============================================================
  Open : /setup
  Token: <one-time-token>
  (one-time — consumed the moment you finish setup)
============================================================

Open shows the bare path /setup by default — open it on whatever host:port you bound the server to. Set MEM_MESH_PUBLIC_URL to have the banner print a full URL (e.g. https://your-host/setup).

The token is also written next to the DB (/app/data/setup_token), so it survives a mid-onboarding restart. Retrieve it any time:

docker exec mem-mesh-prod cat /app/data/setup_token
# or scan the logs
docker compose logs mem-mesh | grep -A1 "Token:"

Open /setup, enter the token plus an admin username (default admin) and password (≥ 8 chars). On submit mem-mesh saves the credential, enables Basic Auth, consumes the token (single-use), and logs you straight into the dashboard. On a fresh server the first page load auto-redirects to /setup.

Once auth is configured the token is deleted on every startup, so a leftover token can never reconfigure an already-secured server.

Reset the token (lost it, and auth is not configured yet) — delete the file and restart; ensure_setup_token() is idempotent, so a plain restart keeps the same value:

docker exec mem-mesh-prod rm -f /app/data/setup_token
docker restart mem-mesh-prod
docker logs mem-mesh-prod 2>&1 | grep -A1 "Token:"

---

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
python -m pytest tests/ -v

# Format and lint
black app/ tests/
ruff check app/ tests/

# Check embedding migration status
python scripts/migrate_embeddings.py --check-only

---

Documentation

• CLAUDE.md — AI tool checklist (MUST/SHOULD/MAY rules, security policy)
• AGENTS.md — Project context, Golden Rules, Context Map, session management details

AI agent rules

File	Purpose
DEFAULT_PROMPT.md	Standalone behavior rules for projects without installed hooks
modules/	Optional Rule Manager modules: core, search, memory-log, pins, relations, batch, security

Generated hook rules share the installed hook prompt version:

mem-mesh hooks rules --project-id <project-id> --format plain
mem-mesh hooks rules --project-id <project-id> --format claude

Architecture docs

• app/core/AGENTS.md — Core service internals
• app/mcp_common/AGENTS.md — MCP common layer

---

Contributing

1. Open an issue or pull request
2. Follow black and ruff formatting
3. Add tests for any new behavior

See CONTRIBUTING.md for details and CHANGELOG.md for release history.

---

License

MIT