A computed trust traffic light
Green, yellow or red — derived from evidence, verification, freshness and dependencies, never picked by hand. Every color ships its color_reason, so you can always audit the why.
COGO
The memory with a trust traffic light — plus the guard that x-rays what a model tells you.
When you build software — yourself or with an AI agent — you accumulate “truths”: “the database lives on that host”, “the bug is caused by X”, “we decided Y”. Over time they rot: some were never verified, some went stale, some were a hunch. The problem is they all look equally true. COGO, part of the Escriba family, stores that knowledge as Markdown notes and gives each one a confidence color it computes itself: green (verified), yellow (probable), red (assumption). You never choose the color — COGO derives it from evidence, verification, freshness and dependencies, so it is auditable and cannot lie.
The other half is Guard: paste any model’s turn and COGO x-rays it against an ontology of 108 manipulation techniques, showing you receipts when the model contradicts what it said before. It is a single Go binary (a ~12 MB scratch Docker image) that is a web viewer, an MCP server and a CLI at once — MCP-first, deterministic at the core, with models as strictly optional accessories.
Who it is for
Section titled “Who it is for”- Anyone building software with AI agents (Claude Code, Cursor, Copilot) who wants the agent to know which project “truths” to trust as facts — and which to re-check before acting.
- Developers tired of watching notes, decisions and debugging hunches rot silently, all looking equally certain three weeks later.
- Teams running several agents on the same project from different machines, who need shared memory and real coordination between them.
- Anyone who chats with LLMs and wants a second pair of eyes against being pushed — the “jailbreak of the human”: fear, urgency, guilt, gaslighting.
- Self-hosters who want a memory that is just Markdown files in a volume: portable, diffable, and shared by every agent that connects.
Key features
Section titled “Key features”A plain Markdown vault
The vault is the single source of truth: portable, diffable, and it outlives the tool. Everything else is a thin client or a rebuildable cache.
Evidence that gets re-checked
A note can cite a file on GitHub (github://owner/repo@main/src/foo.go:42) and COGO anchors the citation to the blob SHA: when the file changes upstream, the note drops to yellow asking to be re-verified. And stash stores artifacts under their SHA-256, so verify recomputes the hash instead of trusting a citation that rots.
Guard: the anti-manipulation x-ray
Reads every model turn with the adversary’s playbook in hand — 108 techniques distilled from persuasion, interrogation, negotiation, coercion, emotional manipulation and rhetoric, each with its source, its signature in a chat, and its countermeasure.
Receipts against gaslighting
Because COGO sees the transcript, when a model denies what it said, COGO finds the turn where it did and shows both quotes side by side. It stops being your word against its.
Veracity x-ray
Paste a model’s answer and COGO x-rays it sentence by sentence, deterministically: commitment (hedged or asserted), evidence (observed, reported, none) and falsifiability. Strong claims with nothing behind them come out red.
Freshness decay
Things expire: a command lasts 30 days, an architecture decision 180. Old notes decay to yellow and then to red on their own — the traffic light always reflects the current state.
MCP-first: 14 tools
pack, search, open, capture, verify, archive, restore, remove, stash, recall, reflect, lease, guard, xray. What Claude learns today, Cursor reads tomorrow — the same vault.
Shared memory across machines
Over MCP on HTTP with a token, several agents on several machines share one vault. recall works as a cursor: hand back the one it gave you and you get only what changed. And lease takes a time-bounded claim on a resource before a migration or a deploy, so two agents don’t trample the same work.
Built-in GitHub explorer
Connect a repository and browse it from the viewer: a classic file tree or a graph map colored by confidence, showing which parts of the code have verified knowledge behind them and which are flying blind.
Colored context packs
pack "<topic>" assembles the notes on a topic into an ordered context block colored by confidence — degrading the red ones so no agent acts on a hunch as if it were a fact.
A ~12 MB binary you can verify
A single static Go binary (scratch Docker image) that is web viewer, MCP server and CLI at once. Every image is signed with cosign and ships SLSA provenance: you can prove it came out of this repository’s CI.
How it compares
Section titled “How it compares”COGO isn’t the first tool to store what you know, and not even the first to verify it.
Notion and Confluence let a person mark a page as verified with an expiry date. GitHub’s
Copilot Memory pins facts to code citations and re-checks them against the current branch.
Several agent-memory projects carry a confidence field.
Three things, however, are genuinely ours:
- A color the model is forbidden to write. Where other tools have a confidence field, it is usually the LLM that fills it in — and once written it never changes again. COGO computes it from the evidence and recomputes it every time you look.
- Doubt that propagates. Everyone else resolves contradictions pairwise and stops there. A survey of 435 papers on agent memory names this as an open problem: “supersession is local; derived records are not re-examined.” COGO’s
min()over the weakest dependency is exactly that missing piece. - Quarantine instead of filtering. Other systems use verification to hide what is doubtful from the agent. COGO hands it over labelled as an assumption, with instructions not to act on it — because hiding it means the agent doesn’t know what it doesn’t know.