Agentmemory

Agentmemory wires twelve auto-capture hooks into your coding agent. Every session start, prompt, tool call, and stop lands in the memory pipeline without glue code. Each event becomes a compressed observation stamped with its origin and the agent that produced it. Install the plugin and capture begins. That's it.

Agent memory recall here isn't one lookup. The primary recall path ranks BM25, vector, and knowledge graph scores together, then reranks on device. That knowledge graph retrieval angle is what catches both exact keyword matches and meaning-based ones. A plain vector store often misses the second kind. Superseded versions stay out of results while their history stays queryable.

Once you supply an LLM provider key, consolidation runs at session stop. Raw observations compress into semantic memories, duplicates merge, stale rows decay with retention scoring, and audit rows record the whole sweep. This is what keeps the store from turning into an unsearchable pile of logs.

This is a full MCP memory server, not a thin wrapper. Fifty-four MCP tools cover saving, recall, smart search, sessions, governance, audit, and export behind a single server. Every one has a REST twin under /agentmemory/*, so you can curl it, fetch it from a browser, or proxy it from your own agent. Saving a near-duplicate returns a similarTo hint instead of a second copy.

With an LLM key set and graph extraction enabled, entities and relations extract from observations, including temporal edges. Every record carries write-time provenance, so you can tell whether a memory came from the user, the agent, a tool, an import, or a shared source.

Lessons live in a dedicated BM25 index with confidence and recency reranking, so a rule you save once comes back when it's relevant. Point the runtime at a Claude Code JSONL transcript and it rehydrates the session, indexes it, and derives crystals and lessons from what it finds.

A local viewer on port 3113 lets you watch observations arrive in real time and inspect the graph your agent has built. It's the quickest way to confirm capture, indexing, and recall are all working before you trust the store with real project memory.