A memory layer for AI agents that outlives your terminal session

What it does

agentmemory is a local memory server that sits between you and your AI coding agents. It remembers context across sessions — architecture decisions, bug fixes, preferences — so you don’t re-explain your codebase every time you open a new chat. It speaks MCP, REST, and native hooks to Claude Code, Cursor, Copilot, Codex, Gemini CLI, and roughly a dozen others.

The interesting bit

The project extends Karpathy’s LLM Wiki pattern with confidence scoring, lifecycle management, and knowledge graphs — then wraps it in a zero-dependency server (no external DB required). One memory store, shared across every agent you use. The “iii engine” underneath handles hybrid search and claims 95.2% retrieval recall at top-5.

Key highlights

• 53 MCP tools and 12 auto-hooks for Claude Code alone; most agents get native plugins or MCP wiring
• Claims 92% fewer tokens spent on re-contextualization per session
• 950+ tests, CI passing, shipped as a single global npm binary (agentmemory on :3111)
• Includes a real-time viewer and console for inspecting what the system actually remembered
• Supports 17+ agents at last count, with more via generic MCP/REST

Caveats

• The README is heavy on integration logos and light on implementation details; the actual storage mechanism and conflict resolution are unclear
• Benchmarks are self-reported (R@5, token savings) with no link to independent evaluation or methodology
• The “0 external DBs” badge is technically true, but leaves open whether this scales beyond single-machine use

Verdict

Worth a look if you juggle multiple AI agents and are tired of groundhog-day explanations. Skip it if you need distributed team memory or want to audit the storage engine before trusting it with proprietary code.