Skip to content

markgatcha/memos

v1.6.27MIT

Local-first persistent memory: store durable facts, recall them in any session. Zero cloud, zero telemetry, SQLite-backed.

GitHubOpen Website

MemOS plugin for Claude Code

Local-first persistent memory for your agent. Installing this plugin registers the MemOS MCP server (21 tools: store, search, retrieve, forget, revert, graph, context packs, temporal search, trust, link, quarantine release, fact extraction, diagnostics, reminders, …), plus the /memos, /recall, /forget, /revert, and /timeline slash commands and the memos-memory skill.

Install

/plugin marketplace add Markgatcha/memos
/plugin install memos@memos-marketplace

The MCP server runs via npx -y @mem-os/sdk mcp and stores everything in ~/.memos/memos.db on your machine — no cloud, no API keys.

Commands

  • /memos <fact> — store a durable fact
  • /recall <query> — recall relevant memories
  • /forget <what> — fuzzy-match a memory, confirm, delete it
  • /revert <what> — revert a memory to its previous version ("that", "the last thing", "what I said about X"), confirm first
  • /timeline [limit] — show what MemOS has learned recently, newest first

Hooks

  • SessionStart — injects a compact summary of remembered context into new sessions.
  • SessionEnd — session wrap-up bookkeeping.
  • UserPromptSubmit (auto-capture) — notices explicit "remember this" / "don't forget" phrasing in your prompts and stores the fact automatically, no command needed. Conservative: only explicit requests trigger it, slash commands are skipped (no double-store with /memos), and anything shaped like a credential is never stored.

Skill

memos-memory teaches the agent when to store (preferences, decisions, environment facts, corrections) and when to recall (memos_context_pack before answering questions that may depend on prior sessions), plus hygiene rules (supersede instead of duplicating, forget only on request).

The skill now also teaches contextual storage: pass context (topic, date, and the previous turn) and metadata.sessionId / timestamp / speaker on memos_store. The MCP manifest sets MEMOS_CONTEXTUAL=1, so the session/date/speaker header is added to the embedding automatically. On LoCoMo retrieval-only this moved Hit@10 from 69.2% to 80.1% and all-evidence recall@10 from 59.6% to 71.5% (same embedder, fusion, top-10). This works identically in Claude Code and Hermes — both consume the same skills/memos-memory/SKILL.md and mcp.json.

Experimental: local reranking

An opt-in cross-encoder reranker re-scores the top fused candidates for sharper retrieval. Default off. Enable with:

{ "experimental": { "rerank": { "provider": "local" } } }

The default model is Xenova/bge-reranker-base (~278MB, downloaded on first use). On a 120-question LongMemEval set it reached hit@5 1.0000 (120/120), evidence recall 0.943, vs 0.958/0.862 for the fused baseline — clearing the accuracy gate on every metric. Override with localModel or MEMOS_RERANK_MODEL. Expect ~60s/question on CPU; a GPU or a /rerank endpoint is recommended for interactive use.