Introduction
kaibo
kaibo is a memory layer between an agent and everything it has seen. You send raw events. kaibo keeps what is current, retires what is stale, and returns the smallest context that answers a task, with the events it came from.
There are four operations. They are the same in the REST API, the TypeScript SDK, the Python SDK, the CLI and the MCP server.
| Operation | What it does |
remember | Store raw events. They are appended to a hash-chained log, then turned into memories. |
recall | Get the smallest context for a question, with a citation for every item. Says so when it has no answer. |
explain | Show where a memory came from: what it replaced and the exact source events. |
forget | Delete by id or query. Recall stops returning it at once. Returns a receipt you can verify. |
Status
Early. Everything above works. A model writes memories from raw events, tags each with the categories it belongs to, and recall searches by meaning, by keyword and through that concept graph, so "what foods do I like" finds "I like apples". If the model is unavailable or over its daily budget, kaibo falls back to deterministic rules and keyword recall. Accuracy is still being measured on RecallBench and the numbers are not published yet. The SDKs, CLI and MCP server are built but not published to a registry yet.
Where to go next
- Quickstart: send an event and read it back in four ways.
- Concepts: how events, memories, recall and forgetting fit together.
- API reference: every endpoint, generated from the same schemas the server validates with.
- MCP: give an agent the four operations as tools.