akitaonrails/ai-memory

The problem

Coding agents now ship their own memory, but it is trapped: one agent, one laptop, one person. Hand a half-finished task from Claude Code to Codex, or resume the same repo on another machine, and you re-explain the architecture, the dead ends, and the open questions. ai-memory puts capture, storage and retrieval behind a shared server so every harness in a project reads the same memory.

What's different

The usual approaches split into fact extractors (Mem0/LangMem-style per-turn facts), temporal knowledge graphs (Zep, Graphiti), and file-first markdown tools (basic-memory). ai-memory picks the file-first side and pushes it further: the source of truth is a git-backed wiki of plain .md pages, with a derived index that can be rebuilt, so nothing is locked into a binary store or vector DB. On top of that it adds capture from lifecycle hooks across a long list of supported harnesses, plus typed, claim-once handoffs between agents. Memory aging — decay, dedup, contradiction flagging — runs with zero LLM calls and is reversible via git version history; the LLM-backed consolidation is opt-in and off by default. Whether that is genuinely better than a per-turn fact store depends on whether you value editable pages over atomic rows.

Who it's for

Developers who regularly switch between coding agents or machines on the same repositories, and small teams that want one shared project memory with per-person attribution. It fits people willing to run a server, on a laptop or homelab. It is a poor fit if you only ever use one agent on one machine and its built-in memory is enough, or if you want a fully hosted zero-ops service — this is self-run infrastructure. Anyone expecting a polished, long-battle-tested product should also wait.

Cost of adoption

You need Rust 1.95+ to build, a server process to run, and per-harness hook/MCP registration (the support matrix is long but uneven: some agents are hooks-only or MCP-only, and native Windows is experimental — WSL2 is the supported path). The zero-LLM default keeps it cheap, but the README is thin on operational details: no benchmark numbers beyond an asserted write ceiling, and the aging/dreaming features require tuning to be useful. MIT licensed, single binary, no paid tier, which lowers lock-in risk. Maturity is the real cost: the project is recent and the docs are the main source of truth.

Verdict

A genuinely differentiated idea — agent-agnostic, file-first, LLM-optional memory — aimed at a real pain point that agent vendors will keep half-solving in incompatible ways. The momentum is real but young: 13.2k stars in about five months, with +3,449 in a single week, and no trending-board or Hacker News footprint to corroborate the spike. That pattern suggests organic interest plus a burst, not yet evidence of a stable user base. Worth trialling if you already switch harnesses and can tolerate a self-hosted server and early-stage docs; otherwise bookmark it and revisit once the operational story settles.