Recurring build failures
WhenAgent sees a Vercel/CI error it hasn't seen before
Check env scoping — Alice's fix from March. Deploy only after NEXT_PUBLIC_ vars match staging.
Fix retrieved in 4s. Previously took 45 min of bisecting.
Loading
Loading
01 · Developers
Last month that was you, in a session it no longer remembers. Next quarter it is the teammate who joined after the fix. Either way it bisects for forty-five minutes and lands somewhere wrong, because nothing told it the answer exists. Cognition is an MCP server that hands it the fix, with the author and the reasoning, before it starts guessing.
$ curl -fsSL https://www.cognitionus.com/api/install | COGNITION_KEY=cog_me_xxx bashScoped key · revocable any time · no source code leaves the session
01 / 07 · Why you need it
02 / 07 · Install
$ curl -fsSL https://www.cognitionus.com/api/install | COGNITION_KEY=cog_me_xxx bashScoped key · revocable any time · no source code leaves the session
Create one at /start. It is scoped to you and to the host you pick, and you can revoke it from the dashboard whenever you want.
The installer writes the MCP entry and a short instruction block for that agent. Nothing else on the machine changes.
Hosts only reload MCP config on a full restart. From then on the agent checks what is already known before it guesses.
Something off after the restart? curl -fsSL https://www.cognitionus.com/api/install/claude-doctor | sh tells you what it sees. Full walkthrough in the docs.
03 / 07 · Compared
If you are one person with rules that fit in one .md file, that file is enough today. It stops being enough the first time you forget to update it, or the third person joins and nobody knows which line to trust. Cognition is what that file cannot be: attributed, approved, and self-correcting.
| .md file | Vector memory | Cognition | |
|---|---|---|---|
| Keeps the author and the reasoning | no | no | yes |
| A human approves before the team sees it | no | no | yes |
| Wrong entries decay instead of misleading the next agent | no | no | yes |
| Follows you across Claude Code, Codex, Cursor, Gemini | copy by hand | depends | yes |
| Reports back whether it helped | no | no | yes |
| Fits in one file, zero setup | yes | no | no |
100%
behavioral cases passed
.md files, RAG, and mem0-class memory scored 0% on the same cases.
86.7%
recall@1
With 0% abstention false positives. Context files hit 93% recall but fired on 60% of off-topic queries.
0%
cross-user leak rate
Personal memory stayed personal in every retrieval case.
Internal behavioral and retrieval tests. Conditions, cases, and what each number does not prove are on the benchmarks page.
04 / 07 · Engineering
The best engineers on your team are debugging the same classes of bugs every month. Cognition captures the fix once, attributes it to the person who found it, and routes every future agent to it before they guess.
WhenAgent sees a Vercel/CI error it hasn't seen before
Check env scoping — Alice's fix from March. Deploy only after NEXT_PUBLIC_ vars match staging.
Fix retrieved in 4s. Previously took 45 min of bisecting.
WhenOAuth callback race condition on mobile
Add 300ms debounce + session lock before redirect. Root cause: token write < redirect.
Two engineers avoided re-diagnosing; one blocked PR caught pre-merge.
WhenPR opens touching the data layer
Never mutate shared state in useEffect. Return cleanup. Enforced since Q1 incident.
New teammates review like seniors. Taste travels with the repo.
WhenCI fails inconsistently on timing-sensitive tests
Mock Date.now() and advance fake timers. Real async causes non-determinism.
Test suite green. Root cause documented, not rediscovered.
05 / 07 · Onboarding
Onboarding is a lossy process. The knowledge that made your team fast lives in Slack history and senior engineers' heads. Cognition makes it retrievable from day one.
WhenNew engineer opens any file
Hooks in /hooks/. State co-located with component unless shared. No barrel exports (causes circular deps). PR title: <type>(<scope>): <subject>.
Day-1 PRs don't need style corrections. Code review is about logic, not formatting.
WhenFirst day, new machine
1Password → Vercel → Supabase → Linear → Slack #engineering-access. AWS requires MFA hardware key. Request in #it-access.
Setup takes 2h instead of 2 days. IT tickets drop 80% in first week.
WhenBug in unfamiliar part of codebase
Auth layer: Alice. Billing: Bob. ML pipeline: Carol (async; she's in Berlin). For urgent: Slack @here in #infra, not DM.
New engineers don't interrupt the wrong person. Response time drops.
WhenTouching a legacy service
payments-v1 uses class components — don't convert without migration plan. Legacy DB still running Postgres 13; no JSON operators. Migration in Q3.
Breaking changes caught before they're written, not in review.
06 / 07 · How it compounds
Each approved skill is retrieved, applied, and scored. What keeps helping gets stronger. What stops helping fades before it can mislead anyone.
07 / 07 · Good questions
Three things. Approved skills: the trigger, the steps, the checks, who wrote it, and when. Outcome receipts: which skill fired, in which agent, and whether it helped. Routing signals that make the next retrieval better. That is the whole list.
No. The agent reads your code the way it always has; Cognition only sees the approved skill. Your source code, secrets, credentials, API keys, names, emails, and unapproved drafts are dropped at the gate and never stored.
No. Cognition returns memory and routing guidance, nothing else. It cannot execute code, modify files, take actions, or run in the background after the session ends. It cannot write to the team brain without a human saying yes.
The agent reports the outcome after each use. A skill that keeps not helping decays and stops being retrieved. You can also say “this memory is wrong” in plain English and the agent routes it to a suppress-and-forget, after you confirm.
Claude Code, Codex, Cursor, and Gemini CLI have first-party installers. Anything that speaks MCP over Streamable HTTP can use the generic installer. Every host needs a restart after install, because that is when hosts reload MCP config.
It adds one authenticated MCP call before meaningful work starts, and a short outcome call at the end. Set that against the forty-five minutes it spends rediscovering a fix it could have been handed.
Revoke the key from the dashboard and the server stops answering. The MCP entry is one block in ~/.claude.json, ~/.codex/config.toml, or ~/.cursor/mcp.json. Personal memory can be deleted from account settings; email privacy@cognitionus.com for full deletion.
$24.99 per seat for individuals and small teams. Teams of five or more pay a $199 flat fee. One avoided forty-five-minute rediscovery a month covers it.
The full boundary, key scopes, and deletion path are on the security page.
Every session without it, the agent starts from zero and you pay for the fix again. One command, one restart, and the first skill you approve is the last time it has to be rediscovered.