your agent re-reads everything,
every session.

The agent asks for what’s relevant and gets exactly that — including the attempt that failed last time. Recall is hybrid: full-text and semantic search fused with learned weight, recency, and verified outcomes. Every recall is traceable — you can ask why a memory surfaced.

> memory_recall("auth token refresh bug")

1. [episode] Tried refreshing tokens in middleware — expected silent
   renewal, got logout loop. Root cause: clock skew. (weight 2.4)
2. [ltm] Auth tokens are minted with 90-day expiry in kairos-token.ts,
   scope "kairos". (weight 1.9)
3. [stm] Hypothesis: refresh failures spike after laptop sleep. (weight 0.8)

install is idempotent — re-running repairs drift. It registers the MCP server for every Claude Code session, keeps the local embedding runtime warm, and opens each session with a recall digest of durable repo knowledge. Other MCP agents connect with backant-memory print-config.

Kairos runs on this exact memory — same store, same namespaces, under ~/.claude/kairos/memory. Use them together and your autonomous developer and your interactive sessions share one brain per repo. Or use memory alone; it does not need Kairos.

Memories are stored per repo, resolved from your git origin. Embeddings are computed locally — a Qwen3 embedding model served by Ollama in Docker on your machine. There are no remote embedding APIs. Ever. Elastic-2.0 licensed, source on GitHub.