// the take
A self-hosted memory server for AI coding agents (Claude Code, Codex, etc.) that sits in front of them as a drop-in proxy, turning conversations, docs, and codebases into four asset types — chat memory, skills, a doc wiki, and a code call-graph — with team-level sharing and ACLs. It's aimed at teams running multiple agents who want shared, persistent context instead of re-explaining the project every session.
The L0-L3 layering (raw conversation → extracted facts → scenario blocks → persona summary) with a BM25 + vector + RRF fallback and explicit item/char/timeout caps is a real answer to the 'memory floods the context window' problem, not just a vector DB wrapper. The proxy-based integration is genuinely low-friction: point an agent's base URL at it and it works without a plugin, hook, or MCP server, which is a meaningfully different integration model than most memory tools that require SDK glue code. Storage is pluggable (SQLite, MongoDB, Tencent VectorDB) behind a common adapter interface, and the ACL model (private/team/restricted/agent-scoped) is a real permission system, not a single-tenant afterthought. CodeGraph and the Skill asset system are built on top of existing open-source work (codegraph, Hermes Agent) rather than reinvented from scratch.
CodeGraph only indexes public HTTPS repos — private repos and SSH auth are explicitly unsupported, which is a significant gap for a tool pitched at internal team codebases, most of which are private. The MongoDB backend is marked experimental and off by default, so anyone past the SQLite/toy-scale tier has to adopt Tencent's own VectorDB to get a production-grade store. The one benchmark offered (PersonaMem 48%→76%) has no methodology, no baseline comparison against other memory systems, and no reproduction steps — that's a thin basis for the main efficacy claim. The README's own notes admit automated memory routing is 'still under iteration,' meaning the core pitch — agents automatically getting the right memory — currently requires manual asset binding. Getting started means running three separate services (memory-core, memory-hub, proxy) plus configuring two separate sets of LLM credentials, a heavier operational footprint than most memory libraries. It's also v2.0.0 with the team-memory feature labeled Beta, so expect rough edges and possible breaking changes despite the large install base.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →