MEMORY
Stable facts
- User preferences
- Project conventions
- Recurring tool quirks
- Decisions that change future behavior
AGENT MEMORY TRIAGE / FIELD NOTE 001
Stop asking the agent to remember everything.
Most “AI forgot” problems are really filing problems. Memory, logs, docs, receipts, and trash are different places. Mixing them turns yesterday’s noise into tomorrow’s instruction.
Five buckets. One decision per note.
Before you paste another summary into memory, decide what kind of object it is. The wrong bucket costs future attention.
SESSION LOG
Handoff state
- What happened today
- What is blocked
- Next concrete move
- Links to evidence
PROJECT DOCS
Shared truth
- Architecture
- APIs
- Runbooks
- Design principles
RECEIPTS
Raw evidence
- Tool output
- URLs
- Commit history
- Published posts
TRASH
Expired noise
- Completed task notes
- Temporary constraints
- Duplicate summaries
- Stale plans
The triage checklist.
Run this before promoting anything into durable memory. If it fails, store it somewhere else or delete it.
Will this still matter next week?If no, it is probably a session log, not memory.
Will it change a future decision?If it only records that work happened, link to the receipt instead.
Is it a preference, convention, or constraint?Those belong in memory when they recur across sessions.
Can it become dangerous if stale?Temporary bans, sprint decisions, and “do not touch X today” notes need expiry.
Can the agent retrieve the raw evidence elsewhere?Do not stuff transcripts, logs, and tool dumps into the prompt. Point to them.
Can it be shorter?Memory should be a sharp handle, not the whole suitcase.
My rule: memory earns rent.
If a note will not protect a future decision, prevent a repeated mistake, or preserve a stable preference, it does not belong in durable memory.
If this is the problem you are fighting in Claude Code or Codex, start with the free Claude Operator Starter. It is the soft entry point, not a cold pitch.