vibehacker
Discuss
Sam Okada
2 months ago

cognee memory skill recalled a deleted .env path from tuesday

Cognee Agent Skills
Persistent long-term memory skills for agents—install, ingest, recall, forget

was demoing agent memory for a Tokyo meetup dry-run. installed the cognee agent skills pack, ran ingest on our CLAUDE.md + a week of session notes.

recall then confidently returned /Users/sam/old-sideproject/.env.local — a path I deleted tuesday after the vercel migrate. the agent tried to cat it mid-demo. screenshots saved me.

forget skill fixed that one entry. still weird that "long-term memory" had no TTL hint in the skill docs. anyone wiring this into cursor agents for real, or am i holding it wrong?

5 comments

Join the discussion

Log in to comment.

  • Mira

    lol yeah. we ingested a whole monorepo notes dump and it kept resurfacing a postgres url from a staging box we killed in august.

    forget works, but i ended up wrapping ingest so anything matching .env / secret / password gets stripped first. felt dumb that the skill didnt warn about that.

    • Devon Hale

      same class of bug we hit with our own rag table. once a secret lands in the index, "delete the file" does nothing.

      i put a zod check before ingest — reject any chunk that matches a known secret pattern, log the miss to slack, keep moving. slower than a blind dump, but i'd rather that than an agent cat'ing dead credentials on a call.

      • Kayla

        yeah the zod gate before ingest is the move. we had a runner agent dump a deleted .env.staging path into slack mid-pr review because someone had checked the notes folder into the memory store.

        i now fail the ingest job if any chunk matches *_URL= or API_KEY. slower. still cheaper than rotating creds at 11pm portland time.

  • Jules Park

    repro'd something close on a cursor agent with cognee skills last week.

    1. ingest a notes dir that once had .env.example with a real DATABASE_URL paste
    2. delete the file
    3. recall still surfaces the url string in a "related context" blob

    forget cleared the entry. no warning on ingest that the chunk looked like a secret. that's the bug for me — not the recall, the silent accept.

    • Noah Kim

      Silent accept on secret-shaped chunks is the real failure mode.

      If your agent memory has no audit trail for what got ingested — path, hash, who triggered it — you're one demo away from leaking a dead credential on a call. Forget is a mop. Ingest policy is the floor.

More like this

View all