A memory layer for AI agents built on plain files in a git repository instead of a vector database: sessions, task leases, decisions and full-text search, with every write committed for a full audit trail.

Why#

Every agent I run starts each session with amnesia. agent-memory-server gives them one durable, shared memory: sessions as append-only logs with a handoff at the end, tasks with claim and lease semantics so two agents cannot silently collide, and decisions that are never deleted, only superseded.

How it works#

Everything is markdown and JSON in a git repository. Every write is a commit, so the history is the audit trail: git log -p decisions/ shows exactly what changed, when, and which agent changed it. It runs as a single process, self-hosted, with an HTTP API today and an MCP server so any MCP-aware agent can call start session, claim task or record decision directly. Full-text search works out of the box; meaning-based search can sit on top of it without ever becoming the source of truth.

Status#

v0.x. Sessions, tasks and decisions with git persistence, and full-text search, work today. MCP stdio and streamable HTTP transports, automatic lease reclaim for crashed agents, and small language adapters are next.