zero dependencies · single binary · Apache-2.0
Your “OK” to an agent becomes the team’s signed memory.Any agent opening the workspace starts knowing what is true, what was decided, and what not to trust yet — and that memory grows with every session.
or: npm install -g @kervo-os/kervo
the loop
01
kervo compile reads the repository — commands, changes, tasks, modules — into a fact skeleton. Same workspace, same bytes, golden-tested in CI. No LLM in this path, ever.
02
Sessions capture durable facts with evidence — how to run things, decisions made in conversation. Each enters the ledger as a trust-labeled proposal with provenance, never as anonymous truth.
03
Verify or retire in kervo review or the dash — one keystroke, reason recorded. Verified memory returns to every future session, any agent, any teammate, for zero tool calls.
trust lifecycle
Facts are extracted deterministically. Everything else — goals, decisions, risks — enters as a labeled proposal with provenance: who said it, how it was verified. Proposals get verified, go stale, and get retired with their reason shown — never silently dropped.
## Known Decisions
[verified — human:jina]
Payments stay on the legacy gateway until Q3 —
the migration branch is intentionally unmerged.
Evidence: standup decision 2026-06-30
[stale — 41 days]
Staging deploys run from ci/deploy.sh
[generated — agent:claude]
Rate limiter TODOs suggest incomplete rollout
→ awaiting human judgment
kervo dash
One command serves a local dashboard over every kervo workspace on your machine — pending proposals, trust states, last activity — with keyboard-first triage that writes each judgment back to that repo’s own ledger. No daemon, no account, no server: it dies with the command, and truth stays per-repo in git.
measured, not claimed
+29.2pp
poisoning protection, kervo vs unmanaged notes
pre-registered bar was ≥20pp — passed. Blind, n=24.
0 / 54
poisoned claims accepted with the artifact
every actual infection (3/3) happened in the unmanaged arm.
5.5 → 9.5
onboarding answers /10 on a real monorepo
after one write-back cycle. Cost unchanged: 1 tool call.
Same repository, three context arms — kervo artifact, labels stripped, unmanaged notes — seeded false decisions, fresh consumer sessions, judges blind to arm and hypothesis. Grades are agent-judged under a pre-registered rubric; a human-grading replication kit is included but has not been run — the limitation is stated, not hidden. Protocols, pre-registrations, and every raw response:kervo-os/experiments.
Two questions, one scan, and every AI session in this repo starts already knowing your project.