agent-memory-system

Persistent repository memory, worklogs, handoffs, and code graph context for AI coding agents.

Agent Memory System is a TypeScript CLI that creates a repository-local memory layer for AI coding agents. It generates reviewable Markdown and JSON artifacts that describe project structure, development workflow, API/storage/security context, static code graph, agent worklogs, and handoffs.

The goal is to help coding agents resume work across sessions and tools without rediscovering the codebase from scratch.

AMS is not a chatbot memory service, vector database, or agent runtime. It is a repo-local context layer for tools such as Codex, Claude Code, Cursor, Gemini CLI, Antigravity, and other coding agents that can read files before acting.

Current Benchmark Snapshot

Maintainer-run self-benchmark on 21 repository tasks. These results are checked into the repo and should be treated as early measurements, not independent validation.

Read the benchmark summary or inspect the raw run data in benchmarks/runs/ and benchmarks/metrics/.

Why This Exists

AI coding agents often start cold. Even when a previous agent already explored the codebase, the next session may repeat the same file traversal, miss prior decisions, or change shared interfaces without checking downstream usage.

AMS makes that context explicit: