Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.backant.io/llms.txt

Use this file to discover all available pages before exploring further.

Kairos is local-first by design. This page is the canonical statement of what data lives where, what leaves your machine, and how to fully wipe local state if you want to.

​
What stays local

Everything to do with memory, dream passes, and policy lives on your machine.
PathContents
~/.claude/kairos/memory/.index.dbThe memory store: lexical index + dense vectors, all in one SQLite file.
~/.claude/kairos/config.jsonEmbedding model tier, dimensionality, Ollama URL.
~/.claude/kairos/models/Docker-managed local model cache.
~/.claude/kairos/dreams/Dream pipeline logs (per-dream timelines).
~/.claude/kairos/eval/Weekly eval reports and replay outputs.
~/.claude/kairos/.session/Per-turn scratch state.
~/.claude/kairos/.state/Daemon state (PID, latest activity, etc.).
<workspace>/.backant.tomlPer-project policy. Committed to your repo.
No codebase content is sent to BackAnt servers for memory operations. Embedding, indexing, and recall all happen on-host.

​
What leaves your machine

Three outbound traffic surfaces:

Claude API (via your Claude Code installation)

The agent’s reasoning calls go to Anthropic via your existing Claude Code setup, using your own Claude credentials. Same privacy surface and policy as any other Claude Code use — Kairos does not add an additional intermediary.

GitHub (via gh CLI)

PR operations, comments, and CI queries go via the gh CLI against the repos you’ve authenticated. Kairos uses your existing GitHub token; no separate credentials are required.

Subscription heartbeat (backant.io)

A small periodic call validates your subscription is active. Sends: an auth token, a heartbeat timestamp. Does not send: codebase content, memory entries, prompt content, or any user data beyond the auth token.

​
What Kairos never sends

To be explicit, none of these ever leave your machine via Kairos:
  • Memory entries (STM or LTM content)
  • Embedding vectors
  • Dream outputs
  • Eval report contents
  • File contents from your codebase that aren’t already part of a Claude API call you would have made anyway
  • .backant.toml content
The Claude API call for reasoning is the only path by which code-derived content reaches a third party — and that’s the same path Claude Code itself uses.

​
What about telemetry?

Kairos does not emit anonymous telemetry. There is no opt-out toggle because there is nothing to opt out of. If you want to verify, the network surface is small and easily observable: 
# Watch what Kairos talks to
lsof -i -P -n -p $(pgrep -f "backant-daemon")
You should see connections to: api.anthropic.com, api.github.com, mcp.backant.io (heartbeat). Plus whatever your Claude Code installation does on its own.

​
Resetting local state

To wipe all Kairos local state and start clean: 
backant stop
rm -rf ~/.claude/kairos/
You’ll need to re-run backant login, backant memory init, and backant start to set up again.

​
Wipe memory only, keep credentials and model cache

backant stop
rm -rf ~/.claude/kairos/memory/.index.db
rm -rf ~/.claude/kairos/dreams/
rm -rf ~/.claude/kairos/eval/
rm -rf ~/.claude/kairos/.session/
rm -rf ~/.claude/kairos/.state/
backant memory init
This avoids re-pulling the Docker image and model (~3 GB).

​
Wipe per-workspace state only

cd /path/to/workspace
rm -rf .session .state
# .backant.toml is preserved
Restarting the daemon will rebuild session state from disk.

​
Working air-gapped

Once the Docker image and embedding model are pulled, the local memory pipeline works offline. The two things that still need network are:
  1. Claude API calls — required for reasoning. Without these, no work happens.
  2. Subscription heartbeat — required at startup. After startup, brief outages are tolerated.
If you need a fully air-gapped deployment, contact support@backant.io — that’s an enterprise scenario with a different shape.