LabscoMCP early access ↗
JubaKitiashviliMCP SERVER · OFFICIAL PROJECT

Context Mem

Memory that survives sessions plus content-aware compression — a local store behind what its README counts as 45+ MCP tools.

by JubaKitiashvili ↗Agent Memory✓ Official

LABSCO SUMMARY

Keeping everything and spending less context are not opposites here.

The usual tradeoff is forget-or-pay; compressing on the way in and retrieving selectively is what breaks it. Two things distinguish it from a notes plugin: explain_decision reconstructs why something was chosen by walking the evidence chain, and the file-claiming tools exist because parallel agents step on each other. Everything stays local — no cloud tier, no premium tools.

What it is

context-mem ingests every tool call, summarises it with a content-aware summariser, and writes it into a navigable markdown vault backed by SQLite. The store is searchable through MCP with hybrid BM25 plus vector search and an optional LLM judge, so a new session starts with what the last one learned.

What you get

  • Core memory — observe stores with auto-summarisation, importance scoring and entity extraction; recall returns verbatim content by filter; search runs hybrid retrieval; ask answers in natural language over the whole store; timeline and stats show what accumulated and what it cost
  • Knowledge graph — save_knowledge with contradiction detection and validity windows, search_knowledge, promote_knowledge to a global cross-project store, global_search, resolve_contradiction, graph_query, add_relationship, graph_neighbors
  • Temporal and intelligence — temporal_query for what was true at a point in time, time_travel to compare two timestamps, explain_decision to walk the evidence chain backwards, predict_loss, generate_story for a PR description, standup update or ADR
  • Sessions and agents — wake_up for a token-budgeted primer, restore_session, handoff_session, plus agent_register, agent_status, claim_files and agent_broadcast for parallel agents
  • 14 summarisers apply different compression by content type; a session of 50 tool outputs went from 365 KB to 3.2 KB, a verified 99.1% saving, with recent high-importance observations kept verbatim and pinned entries never compressed
  • The vault at .context-mem/vault/ is plain markdown — open it in Obsidian or grep it; the SQLite store is the authoritative record

Requirements

Node.js. npm i context-mem && npx context-mem init detects your editor and writes the MCP config — Claude Code needs no restart, Cursor, Windsurf and VS Code do. Manual setup registers npx context-mem serve. For Claude Code, init also writes 8 hooks into .claude/settings.json that inject relevant memories on prompt submission, so you do not call observe by hand. The vault is opt-in in v3.4 via .context-mem/config.json and default-on in v4.0. A local dashboard runs at http://localhost:3141. Registry version 4.0.0. MIT.

Setup effort

One command — npm i context-mem && npx context-mem init