kaiba: shared memory for humans and multiple AIs

The project began from a specific frustration: AI memory features decide on their own what to remember, which can preserve things a person considers unimportant and drop design decisions or measurements they later want to reference. A second problem was structural. Using Claude Code, Codex, OpenCode workers and Antigravity depending on task and cost means knowledge stays trapped inside each AI, so switching agents triggers hand-offs, and a lesson one AI learned by failing can be repeated by another the next day.

The early version was deliberately minimal, offering only remember and recall, first on top of Honcho and later on SQLite, with bge-m3 embeddings and full comparison search at this scale. Notably, no LLM is used to generate, summarize or reinterpret stored text, because a conclusion someone verified would otherwise become an unverified paraphrase. Writing was also restricted: workers generally receive recall only, and the durable facts they find go into a final report for the orchestrator to judge.

Over time the system moved closest to what the developer calls organizational memory. Issues, PRs and code stay in GitHub or shiori, mandatory work rules become Skills, and kusabi enforces process and review. The author notes that even at around 175 stored items, the attention needed to read the top-5 recalled results became the bottleneck before storage scale did, so retiring unnecessary knowledge and avoiding long-term storage of short-lived information were chosen over adding pgvector or ANN.