agentmemory

agentmemory is a local memory server that captures what a coding agent does, compresses it into searchable memory, and injects relevant context at the start of the next session. It targets the failure everyone hits with CLAUDE.md and .cursorrules — a hand-maintained file that caps out at a couple hundred lines and goes stale — and works across roughly a dozen agents at once, since they all talk to the same server.

The design lineage is explicit: the project's design-doc gist extends Karpathy's LLM Wiki pattern with confidence scoring, lifecycle management, knowledge graphs and hybrid search, and the repository is that document implemented. See llm-wiki-as-agent-memory for the broader convergence.

Usage

npm install -g @agentmemory/agentmemory
agentmemory                          # memory server on :3111
agentmemory demo --serve             # boot, seed sample sessions, prove recall, tear down
agentmemory connect claude-code      # wire MCP into an agent
npx skills add rohitg00/agentmemory -y   # 15 skills so the agent knows when to use the tools

Integration depth varies by agent: native plugin plus 12 hooks plus MCP for Claude Code, 22 hooks for OpenCode, 6 hooks for Codex CLI, plain MCP for Cursor, Gemini CLI, Cline and Goose. OpenHuman can use it as its Memory trait backend, so one store can serve a personal assistant and several coding agents at once.

Runs on the iii engine, pinned to v0.11.2 — it will not attach to a different engine version, which is worth knowing if you already run your own. Embeddings are local (all-MiniLM-L6-v2), no API key, no external database. Windows support goes through WSL2; native setup is manual and agentmemory connect doesn't work there.

Benchmarks

The numbers are prominently advertised, so it's worth separating the two corpora.

LongMemEval-S (ICLR 2025, 500 questions) — the public, comparable one:

coding-agent-life-v1 (in-house, 15 sessions) — hybrid scores P@5 0.240 / R@5 1.000 against a grep baseline at 0.227 / 0.967.

That second table is the more informative one, and to the project's credit the README says so itself: the corpus is small and gold-sparse, 0.240 is the P@5 math ceiling for it, and the lift over grep is "recall + temporal, not aggregate precision" — grep misses one of two gold sessions only on the multi-session temporal query. A benchmark table where the honest reading is "we barely beat grep, here's the one query type where it matters" is a better signal about the project than the headline percentages are.

Token-cost claims (~170K tokens/yr vs ~650K for LLM-summarized context) are self-reported modelling rather than measurement — treat accordingly.

Comparison

The README carries a competitor table against mem0, Letta/MemGPT, Khoj, supermemory, mempalace, hippo-memory and built-in CLAUDE.md. Both MemPalace and Hippo have pages here, and the useful distinction is architectural rather than score-based: Hippo models biological consolidation with decay and half-lives, MemPalace uses a method-of-loci hierarchy over verbatim ChromaDB storage, and agentmemory is the one built as a scored wiki with hybrid search over it. agent-memory-components is the frame for comparing them properly — all three differ mainly at the store layer while making similar extractor choices.

Caveats

The presentation is heavily marketed — Trendshift badge, "#1 persistent memory" in the repo description, stat badges, twelve translated READMEs — and 407 open issues against 25.5k stars suggests adoption running ahead of maintenance. The iii engine version pin is a real coupling: your memory store depends on a second project's protocol staying compatible. None of this is disqualifying, but the benchmark honesty and the packaging polish point in opposite directions, and the benchmarks are the part that earned trust.

Apache-2.0, 25.5k★, 2.1k forks, created February 2026. Repo: https://github.com/rohitg00/agentmemory