Labsco
MCP SERVER

Memory Keeper MCP

by mkreyman

Keep what a long working session learned — decisions, cached file state, the commit it went with — so the next session starts where the last one stopped.

Agent MemoryVerified
Summary
The context that dies with a session gets somewhere to live.

Work is saved under keys with categories and channels and read back filtered or diffed against a checkpoint, so "what did we decide about auth" has an answer that is not a scroll. context_prepare_compaction puts the critical items away before a compaction takes them, and a session can be branched to try a second approach and merged back when it works out.

What it is

A session-scoped context store: items saved under a key with a category, priority and channel, checkpoints over the whole session, and 38 tools for getting things back — filtered, searched, diffed against a point in time, or summarised.

What you get
  • Sessions started against a project directory for git tracking, continued from an earlier one, branched to explore an alternative, and merged back under a conflict-resolution rule
  • Items saved with category, priority, channel and a private flag, and batch save, update and delete as single atomic operations
  • Named checkpoints that can carry cached files and git status, restored later by name or ID
  • Critical context saved ahead of a compaction, and old context compressed toward a target size while named categories are preserved
  • Retrieval that narrows: by key, key pattern, category, channel, priority and created-before/after, with sort, limit and offset
  • Search within a session or across all of them, and natural-language search with a similarity threshold and topK
  • A diff of what changed since a timestamp, a checkpoint or a relative time, and a timeline of activity grouped by period
  • Entities and relationships extracted from saved context, related items walked to a depth you set, and visualization data for the graph
  • A git commit created with the context saved alongside it, plus timestamped journal entries with tags
  • Channels listed with counts and activity, and items reassigned between them with a dryRun first
  • Session data exported and imported through the server's own exports directory
Requirements

Npx on your PATH. No account and no key — context stays with the server, and exports are written to its exports directory.

Setup effort

One command — npx mcp-memory-keeper