localmem 0.1.2

Local-first multi-agent memory MCP server with hybrid search, behavioral graphs, knowledge triples, and lifecycle management

localmem

Local-first multi-agent memory MCP server. Persistent storage for LLM agents — hybrid (dense + sparse) vector search, behavioral pattern graphs, temporal knowledge triples, layered wake-up context, lifecycle management, and a read-only browser dashboard. All on-device, no cloud dependencies.

Exposes its functionality over the Model Context Protocol (SSE transport), so any MCP-capable client — Claude Code, Cursor, Continue, custom agents — can read and write memory through a single shared server.

Why

Most "memory" for LLM agents is either a flat key-value store or a single-agent knowledge graph. Real agent systems need more:

• Per-agent namespaces. Each agent's notes, decisions, and observations stay in its own wing. A reserved shared wing carries cross-agent context.
• Hybrid retrieval. Dense embeddings catch semantic matches, sparse BM25 catches exact terms, RRF fuses both. Either signal alone misses too much.
• Behavioral graphs. Some relationships live in entries; others live in the connections between them — co-occurrence, sequence, community structure.
• Temporal knowledge. Facts change. Knowledge triples track validity windows and surface contradictions automatically.
• Graceful forgetting. Three-tier lifecycle (hot → warm summaries → cold compressed archive) keeps the working set fast without losing history.
• Token-aware loading. Layered wake-up context (L0 manifest → L1 critical → L2 scoped search → L3 verbatim) gives an agent ~170 tokens of high-signal context without pulling the whole store.

Quick start

Use a virtualenv. Recent macOS / Homebrew / Debian Python installs are "externally managed" (PEP 668) and pip install localmem directly will refuse. A dedicated venv also keeps the localmem console script on PATH automatically when activated.

# 1. Create a venv (use Python 3.13 — 3.14 + Apple Silicon has a known
#    sentence-transformers shutdown issue, see Known issues below)
python3.13 -m venv ~/.venvs/localmem
source ~/.venvs/localmem/bin/activate

# 2. Install from PyPI
pip install 'localmem[dashboard,analytics]'

# 3. Scaffold a working config + data directory
mkdir -p ~/localmem-data && cd ~/localmem-data
localmem init --wing my_assistant --dashboard

# 4. Start the MCP server (SSE on http://localhost:8781)
localmem -c localmem.yaml serve

Connect any MCP client to http://localhost:8781/sse and the 22 tools below become available.

Dashboard (optional, read-only):

# In another terminal (venv active):
localmem -c localmem.yaml dashboard            # REST + WS on http://localhost:8782

# The prebuilt React frontend isn't shipped on PyPI yet. To get the UI:
git clone https://github.com/jordanaftermidnight/localmem.git ~/localmem-source
cd ~/localmem-source/dashboard
npm install
VITE_API_URL=http://localhost:8782 VITE_WS_URL=ws://localhost:8782/ws npm run build
cd dist && python3 -m http.server 8785

Then open http://localhost:8785.

Headless / always-on (macOS LaunchAgents):

For a server that survives logout, reboot, and crashes, generate + load the LaunchAgents in one command (after the Quick start above is working):

git clone https://github.com/jordanaftermidnight/localmem.git ~/localmem-source 2>/dev/null
python3 ~/localmem-source/deploy/setup-launchd.py --load
launchctl list | grep localmem

Three services come up: com.localmem.serve (:8781), com.localmem.dashboard (:8782), com.localmem.frontend (:8785). All have RunAtLoad=true and KeepAlive=true — they auto-start on login and respawn on crash. Combined with Docker Desktop's "start on login" + --restart unless-stopped on the Qdrant container (see docs/DEPLOY.md when written), the stack runs entirely headless.

Working from source (contributing or pinning a specific commit):

git clone https://github.com/jordanaftermidnight/localmem.git
cd localmem
python3.13 -m venv .venv && source .venv/bin/activate
pip install -e '.[dev,dashboard,analytics]'

MCP tools

Group	Tool	Purpose
Memory (6)	localmem_store, localmem_search, localmem_retrieve, localmem_update, localmem_pin, localmem_unpin	Entry CRUD + hybrid search
Graph (5)	localmem_graph_add_node, localmem_graph_add_edge, localmem_graph_query, localmem_graph_neighbors, localmem_graph_communities	Behavioral pattern graph
Knowledge (3)	localmem_triple_assert, localmem_triple_query, localmem_triple_contradictions	Temporal triples with contradiction detection
System (3)	localmem_wake, localmem_health, localmem_metrics	Layered wake-up + observability
Intelligence (3)	localmem_intel_detect, localmem_intel_alerts, localmem_intel_report	Pattern detection (opt-in via config)
Operations (2)	localmem_prune, localmem_archive	Retention triggers

Storage stack

• Qdrant — embedded by default (path-backed, single-writer). Switch to a remote Qdrant via storage.qdrant_mode: server
  • storage.qdrant_url to unblock live embedding migrations and multi-process writers.
• NetworkX — in-process directed multigraph with multi-hop traversal and Louvain community detection.
• SQLite (WAL) — temporal triples, agent diaries, wing/room taxonomy, importance scoring with time-decay.

Configuration

localmem.yaml at the repo root is the single source of truth. The shipped defaults run locally with zero edits — set wings: to name your agents and you're done. See inline comments for every section. Highlights:

• wings: [list] — your agent namespaces. shared is implicit.
• embedding.model — all-MiniLM-L6-v2 (384d, fast) or BGE-large (1024d, quality). Switch live with localmem migrate-embeddings --to <model>.
• retention.enabled: true — opt in to the three-tier lifecycle.
• dashboard.auth_enabled: true + bearer key for remote dashboard access.
• intelligence.detectors.* — each pattern detector is off until you point it at a specific wing/room (or node selector for the graph cluster detector). Nothing runs you didn't ask for.

Any string value supports ${VAR} or ${VAR:-default} env-var interpolation, so secrets stay out of YAML on disk.

Dashboard

A read-only browser UI under dashboard/ (Vite + React + dockview). 10 panels: Health, Entries, Metrics, Alerts, Graph, Wings/Rooms, Triples, Diaries, Logs, Admin. Pin/unpin and lifecycle triggers live in Admin. Localhost-only by default; flip on bearer auth to expose it remotely.

Observability

• localmem health and localmem_health MCP tool — per-wing entry counts, store connectivity, embedding device, retention worker status.
• localmem_metrics MCP tool — per-tool call counts, p50/p95/p99 latency, error counts (rolling window).
• /metrics Prometheus exposition endpoint on the dashboard sidecar (text/plain; version=0.0.4). See docs/DASHBOARD.md for the metric reference and example scrape config.
• Structured logging (text or JSON) with optional RotatingFileHandler. See docs/LOGGING.md for Loki + Promtail and ELK + Filebeat shipping configs.

Deployment

deploy/ contains installer scripts for the three major platforms — each generates a config dir, sets up a service (systemd / launchd / Scheduled Tasks), and writes an api_key to a perms-restricted env file:

deploy/setup-ubuntu.sh  --auth --qdrant-server http://qdrant:6333
deploy/setup-macos.sh   --auth
deploy/setup-windows.ps1 -Auth

Documentation

• docs/ARCHITECTURE.md — full specification
• docs/DASHBOARD.md — dashboard panels, auth, metrics
• docs/LIFECYCLE.md — retention / consolidation / archive
• docs/MIGRATIONS.md — schema and embedding migrations
• docs/LOGGING.md — log shipping recipes

Project layout

localmem/
├── src/localmem/         # Package source
├── dashboard/            # React + dockview frontend
├── deploy/               # Installers + service units
├── docs/                 # Architecture, dashboard, lifecycle, migrations, logging
├── manifests/            # Per-agent wake-up manifests
├── tests/                # 300+ tests
├── localmem.yaml         # Default configuration
└── pyproject.toml

Known issues

• Python 3.14 + Apple Silicon + sentence-transformers: the loky process pool used by sentence-transformers can crash silently at shutdown on Python 3.14 / arm64 macOS. Python 3.13 and earlier are unaffected. Either use Python 3.13 (verified end-to-end on this build) or switch to the sparse-only retrieval path via embedding.model: "Qdrant/bm25".

License

MIT. See LICENSE.