Cyberbrain Light
Cited, trust-tiered memory for AI coding agents. One small binary, keyword search, nothing to set up.
Your agent forgets everything between sessions. Light gives it notes that survive: plain Markdown in your repository, keyword search over them, and every hit carrying a citation that resolves back to the exact block it came from.
$ cbl recall 'postgres data directory'
1. r2-867ef2a8cd01 r2 pg18-moves-pgdata (100% of top)
# PostgreSQL 18 moves PGDATA
The official `postgres:18` image puts the data directory at
`/var/lib/postgresql/18/docker` instead of `/var/lib/postgresql/data` ...
caveat: search is lexical: it matches the words in the text, not the meaning.The caveat is not an apology, it is the contract. Light searches for words, and it says so on every result, so nobody builds on a search that did not happen.
Measured on 2026-09-06 against a real store of 1,007 notes, on a server without a GPU (AMD EPYC-Milan, 12 vCPU). Times include process start, because that is what a hook actually pays.
Four ways in
cbl <command> |
init, write, scan, recall, forget, status, doctor, all with --json |
cbl hook <event> |
session-start puts rings 0 and 1 into every new Claude Code session |
cbl mcp |
Model Context Protocol over stdio: recall, recall_id, write, status |
cbl install |
writes the hook into a project’s .claude/settings.json, and takes it out again |
What it does not do
Light is defined as much by what is not in it. There is no semantic search (no model, no embedding), no contradiction check, and no compliance subsystem (egress register, audit chain, PII gate, erasure record, obligations). Those are the reason the full Cyberbrain exists.
Light also has no telemetry, no background sync and no network access of any kind. It makes no outbound connections, not as a setting, but because there is no code in it that could.