Skip to main content
← All use cases

​
The scenario

You have solved this before. You know you have, because the error is familiar and vaguely annoying in a specific way. What you cannot remember is what fixed it. Shell history has the commands but not the reasoning, and cannot answer a question. The issue tracker has the ticket if someone filed one. Usually nobody did, because it took fifteen minutes and fifteen minutes does not feel worth a ticket — right up until you spend fifteen minutes on it again. What this recipe shows that no other does: the memory is written by the agent, unprompted, and read before the work rather than after.

​
Setup

MCP plus the Agent Skill. The skill is doing the real work here — it is what makes the agent recall before starting a task and record after finishing one, without you asking either time.

​
What changes

You debug as normal. When it is fixed, the agent writes down what actually worked — not the six things that didn’t. The next time a similar error appears, it checks first. You see:
“You hit this in March. It was the pgbouncer pool, not the gateway — the recovery order matters: redeploy pgbouncer first, then the gateway.”
The value is entirely in the ordering. The same fact surfaced after you debugged it is a curiosity; surfaced before, it is the fifteen minutes back.

​
Evals

This one is measurable, which most memory claims are not:
  1. Work normally for two weeks with the skill installed.
  2. Keep a count of how often a recalled memory arrived before you started versus after you had already solved it.
  3. The second number being high means the recall trigger is too narrow — the agent is remembering, just too late to matter.
  4. Spot-check three recorded fixes against what actually fixed it. An agent recording its first hypothesis rather than the eventual fix is the common failure, and it poisons the pile.

​
Guardrails

Never let it record secrets. Debugging output is full of connection strings, tokens and keys, and an agent pasting freely will store them verbatim. They are then retrievable by anyone with access to the space and will resurface in later prompts.
  • Record the fix, not the transcript. A blow-by-blow of a debugging session is noise that crowds out the answer. One sentence about what was actually wrong beats forty about what wasn’t.
  • A wrong memory is worse than none, because next time it will be recalled with confidence and send you down the path that didn’t work. Correct them when you spot them.
  • MCP tools load at session start. Adding the server mid-session changes nothing until you restart.

​
Where to go next

The team-scale version, where what one person learns the whole repo knows: Memory for your coding agent.