kaeru
Cognitive memory for LLM agents — local-first, and shareable across a team through the cloud.
An agent that returns, recalls, and reshapes.
Most agents forget everything between sessions. kaeru is the memory they keep instead — a typed property graph to think in, holding episodes and hypotheses while the work is live, settled outcomes and references once it isn't. Open a project and the agent knows what was being thought about, can follow the provenance of a decision, and can pull in what the rest of the team has shared.
The name is 蛙 — kaeru, "frog", a homophone of 帰る "to return" and 変える "to change".
It is a facilitator, not an enforcer. The curator API offers around seventy primitives as available tools; the agent and the person decide when to reach for them. The daemon hints, and never blocks.
- Two tiers A working graph for active thinking and an archival one for settled knowledge — the hippocampus and cortex split, with an explicit, logged promotion between them.
- Bi-temporal Assertion and retraction history is native to the substrate. Read any node as it is now or as of a past moment; a conflict invalidates the old version rather than deleting it.
- Reasoning chains The load-bearing path between two nodes saved as a recallable trail — the "why", not just the "what".
- Local-first A single binary with an embedded CozoDB substrate. No server, no network — until you opt into a shared cloud tier for the team.
- Source-available BSL 1.1, not MIT: read it, change it, run it inside your own organization. Offering it to other people as a hosted service is the one right reserved to us.
Talking to it.
kaeru is an MCP server, so any agent that speaks the Model Context Protocol picks up the whole verb set. There is also a rig adapter for agents that embed the store in-process.
# the re-entry ritual, once a project is open initiatives awake --initiative myproject overview --initiative myproject # capture, then connect — # an unlinked node is an island episode "stand is green after the rollback" link "stand is green…" "rollback decision" chain "first symptom" "rollback decision"
Lamantin AI