Agentmemory

agentmemory is a local-first memory server for coding agents. It captures what an agent did during a session through the agent's own hooks, keeps the record on your machine, and hands the relevant context back at the start of the next session. It works with Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, Codex CLI, Hermes, OpenClaw, Pi, OpenCode, and any MCP client, and runs keyless by default: capture and recall need no account, API key, or cloud service.

When a tool-use hook fires, the observation is deduplicated by hash, passed through a privacy filter that strips secrets and API keys, stored raw, compressed (synthetically by default, or by an LLM when a provider is configured and the option is turned on), embedded when an embedding provider is active, and indexed. At session end the session is summarized, with optional knowledge-graph extraction. At session start the project profile is loaded, a hybrid search runs, and the result is injected into the conversation within a token budget (2,000 tokens by default). Hooks also capture user prompts, file access, tool failures, and sub-agent lifecycle events, and re-inject memory before context compaction.

Memory is consolidated in four tiers: working (raw observations), episodic (compressed session summaries), semantic (extracted facts and patterns), and procedural (workflows and decision patterns). Memories decay on an Ebbinghaus curve, strengthen when accessed, and are evicted when stale, and contradictions are detected and resolved. Retrieval fuses three streams with Reciprocal Rank Fusion: BM25 keyword matching with stemming and synonym expansion (always on), cosine similarity over dense embeddings (when an embedding provider is configured), and knowledge-graph traversal (when entities are detected in the query), diversified to at most three results per session.

Benchmark	agentmemory	Comparison
LongMemEval-S R@5 (500 questions)	95.2%	BM25-only fallback 86.2%
LongMemEval-S R@10	98.6%	BM25-only fallback 94.6%
LongMemEval-S MRR	88.2%	BM25-only fallback 71.5%
Context tokens at 240 observations	~1,900	22K+ when a 200-line memory file is loaded whole
Tokens per year (estimated)	~170K	~650K for LLM-summarized context

The server exposes the same memory over MCP and REST, so several agents can share it: 54 MCP tools (trimmable to a core set of 8), 6 resources, 3 prompts, and 17 skills in the Claude Code SKILL.md format. Multi-agent setups tag every write with an agent role and run in shared or isolated mode, with team namespaces, action leases, signals, and routines for coordination. Every observation carries an immutable origin channel, any memory can be traced back to its source observations, and memory state can be versioned, rolled back, and diffed as git snapshots. A real-time viewer streams new memories and audit entries and replays any recorded session as a timeline.

The runtime is built on the iii engine, which supplies HTTP triggers, key-value state with an in-memory vector index, WebSocket streams, and worker supervision in place of a separate web framework, database, and process manager. Optional LLM-backed compression is auto-detected from the configured provider: Anthropic, MiniMax (Anthropic-compatible), Gemini, OpenRouter, OpenAI, or any OpenAI-compatible local server such as Ollama, LM Studio, vLLM, or llama.cpp. Embeddings default to a local all-MiniLM-L6-v2 model once opted in, with remote providers as an option. Measured workload data puts DeepSeek V4 and Qwen3-Coder within rounding error of Claude Sonnet on the compression task at 10–70 times lower cost.