Home

dm-lite

daimon-memory v2: a small typed memory engine for AI agents. One binary, client/server, hybrid recall, bitemporal history, and multitenant storage (one database per tenant).

The README is the quickstart front door. This wiki is the detailed reference.

Pages

• Install and first run
• Save, recall, and time travel
• The graph
• Wiring your agents
• Run as a server
• Run as a client
• Persona and governance
• Embedding models
• Migrate, upgrade, and build from source

The mental model

dm-lite is client/server. You run dmem serve (locally as a managed loopback daemon, or on a host) and point clients at it. The CLI, the agent hooks, and the MCP server are all thin clients of that daemon; local versus remote is only the URL. On one machine the server and client live together (the quickstart); split them across hosts when you want shared memory.

It is one memory shared across your AI tools (Claude Code, Codex, Hermes, Devin, OpenCode, Grok CLI, Claude Desktop), with the integration included rather than left to you to wire. Switch tools, keep the same brain. Any MCP client can connect, and receives the persona and protocols too, not just the tools. Several agents can even share one tenant while each keeps its own persona and write attribution (per-agent identity tokens, see "Run as a server").

Memory is typed and curated (the agent saves distilled records with a kind, not raw transcripts), and recall fuses keyword and dense-vector search. It is bitemporal: it tracks both when a fact is true in the world and when you recorded it, so you can recall what was true at a past moment, or what you knew at one. Nothing is overwritten. Records can also be linked into a lightweight graph, with recurring domain nouns minted as typed entities, so a recall pulls in a hit's neighbours rather than one isolated note.