LabscoMCP early access ↗
roboticforceMCP SERVER · OFFICIAL PROJECT

Sugar

A local memory store your coding agent writes to and reads from, scoped per project and across projects.

by roboticforce ↗Agent Memory✓ Official

LABSCO SUMMARY

Two stores, and the split is the design.

Most memory tools give you one bucket, which means either your project notes drown your general standards or the reverse. Reserving result slots for guidelines from the global store fixes that directly: a brand-new project inherits how you work, and a mature one still gets reminded of it. The per-type expiry is the other quiet good idea — an error pattern from last quarter should age out, an architecture decision should not.

What it is

A memory layer for AI coding agents that lives in SQLite on your disk. Two stores — one inside the project, one global — and a search that reads both, so project context wins while your cross-project standards still surface.

What you get

  • search_memory — searches both stores and labels which scope each result came from
  • store_learning — saves a memory; passing scope: "global" writes to the cross-project store
  • recall — formatted markdown context for a topic, ready to read
  • get_project_context — a project summary including the global guidelines that apply
  • list_recent_memories — browse recent memories filtered by type
  • Seven memory types with different lifetimes: decision, preference, file_context and guideline never expire, while error_pattern lasts 90 days, research 60 and outcome 30
  • A search that reserves slots for global guidelines, so a mature project's local context cannot bury your standards
  • Resources sugar://project/context, sugar://preferences and sugar://global/guidelines
  • An optional second server, sugar mcp tasks, backing a task queue that reads the same memory

Requirements

pipx install sugarai, then sugar init in the project. Memory lives in .sugar/memory.db for the project and ~/.sugar/memory.db globally. Semantic search comes from the [memory] extra — sentence-transformers all-MiniLM-L6-v2, 384-dimension vectors, with sqlite-vec — and without it search falls back to FTS5 keyword matching and then LIKE. The memory server needs no API key and makes no external calls. The task server takes optional ANTHROPIC_API_KEY, GITHUB_TOKEN and SUGAR_DEFAULT_REPO.

Setup effort

One command — pipx install sugarai

See installation options