Most coding-agent setups repeat the same onboarding and bug-handoff work every session; this project treats that repetition as data. agentmemory silently records tool calls and session events, compresses them into structured observations, indexes them with BM25+embeddings+graph fusion, and injects the right context when a new session starts — so agents resume work without re-teaching.
What Sets It Apart
- Hybrid retrieval pipeline: fuses BM25 keyword matching, vector embeddings, and a lightweight knowledge graph with Reciprocal Rank Fusion (RRF). So what: it balances precision on exact file/term matches with semantic recall for paraphrased developer intent.
- Zero-touch capture + MCP tooling: hooks for SessionStart/Stop and PostToolUse plus a 51-tool MCP surface (memory_recall, memory_smart_search, memory_save, memory_sessions, etc.). So what: any MCP-capable agent can auto-share memories without manual add() calls.
- Memory lifecycle & cost controls: four-tier consolidation (working → episodic → semantic → procedural), decay/auto-forget, and token-budgeted session injection. So what: memory grows sustainably and keeps injected context small and relevant, cutting repeated-context token spend dramatically.
- Local-first architecture: ships with SQLite/iii-engine-based local runtime, local embedding option (all-MiniLM-L6-v2) and optional cloud providers. So what: teams can self-host with low cost or opt into managed models for higher-quality compression.
Who it's for and trade-offs
Great fit if you run coding agents across sessions and want them to remember file edits, tests, auth decisions, and recurring patterns without manual notes. It’s specifically useful for developer workflows (local CLI agents, IDE integrations, or multi-agent MCP setups). Look elsewhere if you need a cloud-hosted multi-tenant SaaS with enterprise SLAs out of the box — agentmemory is opinionated toward self-hosting, local embeddings, and tight developer control, and enabling full LLM-backed auto-compression can incur token costs.
Where it fits
Pairs with Claude Code, Cursor, Gemini CLI, Codex CLI, OpenCode and any MCP-capable client; also usable as a standalone REST memory service. Common use cases: session handoff, automated TODO extraction, RAG-style codebase recall, and observability into agent actions via the real-time viewer and iii-console traces.
