AI Memory

AI Memory is a long-term memory server for coding agents. It keeps one shared store of what a project's sessions learned, so you can stop mid-task in one agent, open another in the same directory, and continue without restating the architecture, the failed approaches, or the open questions.

The store is a git-backed wiki of plain Markdown files, with a SQLite index derived from those files and rebuildable at any time. Lifecycle hooks record prompts, tool calls, and session boundaries, sanitised at a typed privacy boundary before storage; at session end the observations are compiled into wiki pages, and the next session receives a bounded brief plus search over full text, entities, links, and optional vectors. Handoffs between agents are typed, owned, and claimed exactly once.

A single Rust binary runs the MCP and HTTP server and owns one data directory holding the wiki, raw transcript segments, SQLite indexes, and logs. Running the server on a reachable host lets several machines and several people share one project memory, with per-person attribution, multi-user auth, and an audit log. LLM providers are optional: capture, search, and handoffs run with no API key, and adding a provider upgrades session summaries and enables semantic search.

Features

• Cross-agent handoff: more than twenty coding agents feed one shared memory, and a handoff carries where you left off, what failed, and what is open
• Markdown source of truth: memory is ordinary .md files you can grep, edit, or sync, with the database as a rebuildable index
• Ambient capture: lifecycle hooks record prompts, tool calls, and session boundaries without a "remember this" step
• Zero-LLM default path: capture, search, and handoffs work with no API key; a provider adds summaries and semantic search
• Multi-machine and team use: one server serves several machines and teammates, with shared project knowledge and personal handoffs kept personal
• Managed workstreams: ai-memory run adds cross-harness session continuity on top of shared memory
• Retrieval fusion: FTS5, entity-match and graph-neighbour ranking, optional vector ranking, and bounded raw-observation fallback are fused into one ranking
• Security ladder: loopback-only by default, then bearer tokens, per-user accounts, OIDC device auth for hooks, and TLS through a reverse proxy
• Read-only web view: starting with --enable-web serves a browser view of the wiki and a JSON API under /api/v1