One task informs the next.
Keep the evidence, propose a change, check it, then reuse what holds up. We improve work context and methods, not the model’s weights.
Memory
Before
A reporting preference has to be repeated in every task.
After the change
A reviewed memory can supply that preference to a later task.
- Read the work record
- Propose a memory change
- Check the difference and reason
- Apply within the allowed scope
- Reuse or reverse if appropriate
What we keep
Before text, proposed text, reason, application state and revision.
How to check the difference
Run the same task with and without memory. Check requirement coverage and incorrect carryover.
Harness
Before
A failed approach repeats without a clear account of why.
After the change
A tested change to the method can guide the next execution.
- Collect execution evidence
- Identify the failure pattern
- Propose a configuration change
- Compare under equal conditions
- Keep the version that passes
What we keep
Task, model, configuration version, failures, outputs and measured cost.
How to check the difference
A development target: repeat matched tasks and include failures, retries, latency and cost.
LabChin research
Before
A useful finding stays buried in one report.
After the change
The finding, its evidence and unresolved questions inform the next investigation.
- Keep the original references
- Separate findings from assumptions
- Record unresolved questions
- Review the next research plan
- Connect the next result
What we keep
Reference versions, observations, interpretations, open questions and the next output.
How to check the difference
A research workflow in development. Check that every interpretation can be traced back to the referenced material.
Memory proposals and conditional reversal are implemented. The complete comparison and promotion loop is still in development. Better results must be demonstrated through matched evaluations.