memorymaster 3.18.0

Production-grade memory reliability system for AI coding agents. Lifecycle-managed claims with citations, conflict detection, steward governance, and MCP integration.

MemoryMaster

Production-grade memory reliability system for AI coding agents.

Lifecycle-managed claims with citations, conflict detection, steward governance, hybrid retrieval, and MCP integration. Give your AI agents persistent, trustworthy memory.

MemoryMaster prevents the #1 problem with agent memory: drift, stale assumptions, and unsafe disclosure. It gives Claude Code, Codex, and any MCP-compatible agent persistent, verifiable memory with a full claim lifecycle, citation tracking, conflict detection, and human-in-the-loop governance.

---

Architecture

MemoryMaster is layered around MCP/CLI entry points, the MemoryService facade, SQLite/Postgres storage, optional Qdrant vector search, scheduled jobs, and the Obsidian wiki/vault layer. The canonical ingest path is:

MCP/CLI -> sensitivity filter -> MemoryService.ingest -> store write -> FTS5 index

The query path is:

query_memory -> MemoryService.query -> storage reads + optional Qdrant candidates -> ranked context

See docs/architecture.md for the current module map, data-flow details, recent PR status, and sensitivity-filter invariants.

Key features

• 6-state lifecycle: candidate → confirmed → stale → superseded → conflicted → archived
• Citation tracking with provenance for every claim
• Hybrid retrieval: vector (sentence-transformers / Gemini) + FTS5 + freshness + confidence
• Context optimizer: query_for_context(budget=4000) returns auto-curated memory that fits your token budget
• Entity graph with typed relationships and alias resolution
• Steward governance: multi-probe validators (filesystem, format, citation, semantic, tool) with proposal review
• Conflict resolution: 5-tier auto (confidence > freshness > citations > LLM > manual)
• Auto-redaction at ingest: JWT, GitHub tokens, Bearer, AWS keys, SSH keys, custom patterns
• LLM Wiki: compiled-truth + append-only timeline articles with progressive-disclosure frontmatter, explored: true|false operator-review marker, and inline > [!contradiction] Obsidian callouts
• Atlas Inbox V1 (new in v3.13.0): WhatsApp ingestion → source/evidence/action proposal lifecycle → Super-Productivity export. Versioned API/CLI contract for downstream consumers (LifeAgent, etc.) — see docs/atlas-api-contract-v1.md. Real provider adapters (OpenAIWhisperTranscriptionProvider, TesseractOcrProvider) behind Protocols; mock providers stay default.
• Dual backend: SQLite (zero-config) and Postgres (full feature parity with pgvector)
• Dream Bridge for bidirectional sync with Claude Code's Auto Dream
• 7-hook stack: recall, classify, validate-wiki, session-start, auto-ingest, precompact, steward-cron

Full feature index lives in docs/handbook.md.

Benchmarks

LongMemEval-S (N=500, retrieval-only) — v3.15.0 now leads the publicly-reported numbers from agentmemory on R@5 and MRR, after wiring sentence-transformers/all-MiniLM-L6-v2 into the bench harness (the v3.14 baseline was unintentionally BM25-only).

Metric	v3.14.0	v3.15.0	agentmemory	Δ vs agentmemory
Recall@5	0.894	0.966	0.952	+0.014 ★
Recall@10	0.942	0.984	0.986	-0.002
MRR	0.799	0.902	0.882	+0.020 ★

Reproduce: python tests/bench_longmemeval.py --retrieval-only. Full methodology, experiment-by-experiment deltas (1 KEEP, 2 REVERT, 3 NULL), and the architectural findings that surfaced along the way live in docs/longmemeval-results.md and docs/v315-experiments/. QA-accuracy pass (with judge) is deferred until provider quotas allow.

Prerequisites

Required (the package won't function without these)

• Python 3.10+ with pip
• Claude Code, Codex, or any MCP-compatible agent
• An LLM provider — pick one: Claude Code OAuth (free if you're a subscriber, set MEMORYMASTER_LLM_PROVIDER=claude_cli), a free Gemini API key from aistudio.google.com, OpenAI, Anthropic API, or local Ollama. The steward, auto-ingest, and wiki-absorb cycles all need an LLM — without one, claims pile up as candidate and never get validated, deduped, or compiled into the wiki.

Strongly recommended (you'll lose ~80% of the value without these)

• Node.js 18+ for graphify and GitNexus — these are the cached intelligence layers that make MemoryMaster cheap to query. Without them, every "what does this codebase do?" question burns tokens cold-exploring files the graph already mapped. The intelligence-first workflow in CLAUDE.md assumes both are installed.
• Obsidian 1.6+ with the Bases core plugin — the wiki engine writes plain Markdown so any editor works, but Obsidian's backlinks, graph view, and Bases dashboards are how you actually navigate wiki-absorb output. Without Obsidian, the wiki is just a folder of files.

Optional (nice to have)

• Docker for Qdrant — vector retrieval. SQLite FTS5 is the default and works out of the box; add Qdrant when you want semantic recall on top of keyword search.

Quick start

pip install "memorymaster[mcp]"
memorymaster --db memorymaster.db init-db
memorymaster-setup     # interactive: hooks, MCP, steward cron, CLAUDE.md / AGENTS.md

That's enough to use the CLI, the MCP server, and the auto-ingest Stop hook.

# Ingest a claim with citation
memorymaster --db memorymaster.db ingest \
  --text "Server uses PostgreSQL 16" \
  --source "session://chat|turn-3|user confirmed"

# Query memory (hybrid retrieval)
memorymaster --db memorymaster.db query "database version" --retrieval-mode hybrid

# Context optimizer — the killer feature for agents
memorymaster --db memorymaster.db context "auth patterns" --budget 4000 --format xml

# Run validation cycle
memorymaster --db memorymaster.db run-cycle

For the one-prompt agent install (paste into any agent with shell access), see docs/handbook.md#one-prompt-agent-install.

Pick your LLM provider

Provider	Env vars	Default model	Cost
Claude Code OAuth (recommended for subscribers)	MEMORYMASTER_LLM_PROVIDER=claude_cli (requires claude CLI on PATH)	claude-haiku-4-5-20251001	included in Claude Code plan
Google Gemini (default)	MEMORYMASTER_LLM_PROVIDER=google + GEMINI_API_KEY=...	gemini-3.1-flash-lite-preview	~free
OpenAI	MEMORYMASTER_LLM_PROVIDER=openai + OPENAI_API_KEY=...	gpt-4o-mini	~$0.001/call
Anthropic API	MEMORYMASTER_LLM_PROVIDER=anthropic + ANTHROPIC_API_KEY=...	claude-haiku-4-5-20251001	~$0.001/call
Ollama (local)	MEMORYMASTER_LLM_PROVIDER=ollama + OLLAMA_URL=http://localhost:11434	llama3.2:3b	free

The claude_cli provider shells out to your local claude --print binary, so it inherits the OAuth session you're already logged into in Claude Code — no API key, no rotator, no quota juggling. Caveat: cold-start adds 3-15s per call (subprocess spawn), so it's ideal for batched/cron paths (steward, wiki-absorb) and not for latency-sensitive recall. Override with MEMORYMASTER_CLAUDE_CLI_BIN and MEMORYMASTER_CLAUDE_CLI_TIMEOUT. On VM installs the OAuth token expires ~24h, so pair with MEMORYMASTER_LLM_FALLBACK_PROVIDER=ollama; desktop tokens don't expire.

For zero-cost offline use, install Ollama, ollama pull llama3.2:3b, and set MEMORYMASTER_LLM_PROVIDER=ollama.

MCP server

{
  "mcpServers": {
    "memorymaster": {
      "command": "memorymaster-mcp",
      "env": {
        "MEMORYMASTER_DEFAULT_DB": "/path/to/memorymaster.db",
        "MEMORYMASTER_WORKSPACE": "/path/to/your/project"
      }
    }
  }
}

22 MCP tools: init_db, ingest_claim, run_cycle, run_steward, classify_query, query_memory, query_for_context, list_claims, redact_claim_payload, pin_claim, compact_memory, list_events, search_verbatim, open_dashboard, list_steward_proposals, resolve_steward_proposal, extract_entities, entity_stats, find_related_claims, quality_scores, recompute_tiers, federated_query.

See .mcp.json.example for the full template.

Backends

Backend	Install	Use case
SQLite	Built-in	Local development, single-agent, zero-config
Postgres	pip install "memorymaster[postgres]"	Team deployment, multi-agent, pgvector search

Docker Compose

Run the full stack (MemoryMaster + Qdrant + Ollama) with one command:

docker compose up -d

See INSTALLATION.md for Kubernetes / Helm.

Development

# Install with dev dependencies
pip install -e ".[dev,mcp,security,embeddings,qdrant]"

# Run tests
pytest tests/ -q

# Lint and format
ruff check memorymaster/ && ruff format memorymaster/

# Performance benchmarks
python benchmarks/perf_smoke.py

See CONTRIBUTING.md for the full workflow.

Documentation

Document	Description
docs/handbook.md	Full operator handbook — hooks, dashboard, steward, dream bridge, troubleshooting, one-prompt install
INSTALLATION.md	Setup guide: pip, Docker, Helm, MCP config
CONTRIBUTING.md	Dev setup, testing, PR workflow
ARCHITECTURE.md	System design and subsystem details
USER_GUIDE.md	Usage, MCP integration, troubleshooting
CHANGELOG.md	Version history and release notes
ROADMAP.md	Release plan and future tracks
docs/enabling-v2-systems.md	v3 statistical classifier + cadence policy opt-in

License

MIT — Built by wolverin0