Overview
agentmemory helps coding agents carry decisions and discoveries across sessions without replaying an ever-growing transcript. Its local-first iii-engine runtime captures lifecycle events, filters and attributes them, turns them into searchable observations, and combines lexical, semantic, and graph signals within scope and token bounds. Later stages can promote, supersede, reinforce, decay, or forget derived knowledge, while multiple hosts and agents use separate interfaces and coordination controls. The system improves governed continuity, not truth: capture is best effort, optional intelligence adds cost, and local operation still requires security and inspection.
Learning outcomes
- Useful agent memory begins with lifecycle capture, privacy filtering, provenance, and durable state rather than a larger prompt or vector index alone.
- BM25 provides the keyless lexical floor; vectors and graphs add conditional paraphrase and relationship reach, but scores do not prove truth.
- Current memory must be able to change: supersession removes old versions from search, while consolidation and retention shape what remains.
- Shared use requires separate controls for recall scope, deliberate sharing, work ownership, signals, audit, and network exposure.
- agentmemory fits local-first coding continuity with an inspectable service; it does not provide hosted SaaS, enterprise identity, or guaranteed truth.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What problem does agentmemory solve for coding agents?
agentmemory carries selected decisions, discoveries, and failure context across coding sessions without replaying the full transcript. It preserves continuity while keeping retrieval bounded.
How does agentmemory turn a tool event into searchable memory?
agentmemory validates and sanitizes the lifecycle event, adds provenance, stores a raw observation, and creates a compressed searchable representation. BM25 is available by default; vectors and graphs are optional.
Does agentmemory retrieval prove that a memory is true or current?
No. Retrieval ranks stored evidence; it does not verify truth or currentness. Supersession, provenance, inspection, and deliberate correction help distinguish current knowledge from history, but derived summaries and relationships remain heuristic.
Is agentmemory suitable for multi-agent or enterprise deployment?
It can support multi-agent continuity with agent scope, team sharing, leases, signals, checkpoints, and audit. It is not a hosted SaaS or a complete enterprise boundary: RBAC, SSO, immutable compliance storage, and guaranteed capture are not implied.