Read this as What reasoning should survive into the next session?
- Failure Trap
- Appending every reflection directly into hot memory until the prefix bloats and the 80 percent ceiling holds.
- Decision Rule
- Separate reflector from curator; promote only generalized, reusable patterns into durable memory.
A session produces more than code
The working agent leaves behind outputs, decisions, rejected alternatives, and constraint chains.
- Artifacts are easy to keep.
- Reasoning evaporates fastest.
- The next session pays to rediscover it.
The reflector extracts candidates
A reflector reads the session and asks what would help the next similar task.
- It emits candidate patterns.
- It captures rejected alternatives.
- It is not allowed to promote directly.
The curator gates promotion
The curator edits, generalizes, and rejects candidates before anything enters durable memory.
- Promotion costs prefix budget.
- Not every observation generalizes.
- The gate prevents memory sludge.
Durable memory gets the useful pattern
Promoted context becomes a named file, note, or spec the future agent can discover and read.
- Filesystem semantics make it addressable.
- Hot memory stays small.
- Warm and cold tiers hold larger context.
The next session inherits the upgrade
Session N+1 starts with better context because the prior reasoning was curated into the substrate.
- The model did not change.
- The harness changed what it could remember.
- The loop closes back into the prefix.
Collapse reproduces the ceiling
If reflection and curation collapse into one append-only step, hot memory bloats and the 80 percent continuity ceiling returns.
- Uncurated notes crowd the prefix.
- Stale context becomes instruction debt.
- The gate is the memory system.