Labsco
MCP SERVER

Agents Remember

by Foxfire1st

Project memory a coding agent reaches by file path — and that gets re-checked against Git before it is trusted.

Agent Memory
Summary
It solves the moment-of-edit problem: the rule shows up where the change is being made, not in a file at the top of the repo.

A top-level instruction file is read once and forgotten by the time the agent is deep in a function. Putting the note at a mirror path means it is reachable from the file itself, with no search step to get wrong. The drift checking is what keeps that honest — a note whose source has moved on is flagged rather than trusted, which is the failure mode every hand-written memory system eventually hits. The cost is process: this expects worktrees, gates and a closeout step, so it fits teams willing to adopt a workflow rather than just install a server.

What it is

A Python server that keeps versioned Markdown notes beside your code at deterministic mirror paths, verifies them against source commits before use, and updates them only after approved work lands.

What you get
  • Path-addressed notes: a source file's context lives at a predictable mirror path, so an agent holding the file reaches the right note without searching or ranking
  • context_packet — the orienting read at the start of a task
  • memory_quality_check and drift_check — notes tested against source commits and fingerprints before they are believed
  • memory_init and memory_baseline_status — a repository brought under memory management, and its baseline state reported
  • resolve_context, for pulling the notes relevant to what is being changed
  • Isolated dual worktrees through worktree_start, worktree_status, worktree_cleanup and worktree_abandon, so memory and code land together or not at all
  • gate_decide and gate_list, for the approval points in the lifecycle
  • dispatch_agent and the parent/child messaging tools, for coordinating more than one agent
  • Repo-owned behaviour in system files — path rules, tools, coding guidelines, branch policy, reporting shape — so the same rules load whichever harness you use
  • Starter packages for Claude Code, Codex, Cursor, Antigravity, VS Code Copilot, Hermes, Pi.dev and OpenClaw
Requirements

Uv for uvx, or pip, with Python 3.11 or higher, and Git configured with a user name and email. Memory lives in the repository under ar-memory by default; teams wanting separate memory repositories use the external layout instead. No account and no key — everything is files and Git.

Setup effort

One command — uvx agents-remember-mcp@latest --config /absolute/path/to/agents-remember-settings.json