MEMORY & LEARNING
Keep the notes.
Learn from the mistakes.
A place for notes, sources, and decisions, so I can pick up a project without starting from scratch.
Here, “learning” means keeping useful corrections and improving the next attempt.
THE IDEA
Keep the useful parts together.
The design uses Obsidian for project notebooks, Git to keep earlier versions, and separate backups.
The source.
Save where an answer came from, so it can be checked later.
The decision.
Record what was chosen and why, including questions that are still open.
The correction.
Write down what went wrong. Improve the checklist or template when that helps.
CONNECTED TO RESEARCH OS
Check what we already know.
Before fresh research, a memory preflight checks the saved, reviewed notes. It helps decide what can be reused, what needs checking again, and what is missing. New findings go through review before joining those notes.
Recorded on 6 August 2026: the build notes describe two research documents approved for memory. A preflight read one and asked for a refresh because a supporting claim was still marked as needing verification.
Those records still list reuse without new research as untested on a real run. This is a saved example, not a fresh check of the whole installation.
Source review, 1 October 2026: the newer research package includes a memory check for operator review: reuse saved work, refresh it, research the gaps, or leave unsuitable material out. Its behavior with real notes on the running setup still needs acceptance testing. A “reuse” suggestion is not proof that those notes answer the whole question.
1 OCTOBER 2026 / WORK IN PROGRESS
The rules need review too.
The latest document review found two Prompt OS guides giving a different order for checking the same research return. It also found a diagnostic that could mistake a model’s refusal for stale instructions.
A correction proposal is prepared, but has not been adopted. Writing down a better rule does not change the running validator or prove the handoff works. The original packet’s matching raw validator result is still missing.
8 OCTOBER 2026 / A CORRECTION IN PRACTICE
Keep both outcomes in the record.
A trial to pass morning project context to Muse failed acceptance. A separate, narrowly corrected recovery then restored the original context and service, according to my command-line receipt. It also reported both temporary permissions disabled and the audit record retained.
The later morning drafts missed that recovery receipt and described recovery as still unverified. Those drafts were corrected before publication. The correction kept the unsuccessful trial and the successful recovery separate, with the source and limits of each result intact.
This is the kind of work memory needs to support: bring a missed result back into the account without turning a repair into proof that the original task succeeded. The automated feed remains off, and another fix or test needs fresh approval.
27 JULY 2026 / FROM THE INCIDENT LOG
One review changed the answer.
A research report picked a winner, but its comparison used different standards for different options.
The support looked stronger than it was.
The preferred option leaned on a company’s promotional report. Articles repeating it were counted as extra support.
The comparison was uneven.
Other options were marked down for similar evidence. Reading the report exposed that mismatch.
The recommendation changed.
The notes record a revised recommendation and clearer rules for comparing sources. They also say demand needed testing before building.
This is a summary of the July 27 incident note. It records a correction, not proof that the revised plan was adopted.
What the saved records show
A separate research result dated July 21, 2026 has been recovered. Its own “audit passed” label is not an independent validator result. The matching original validator output is still missing.
A July 26 incident note describes automated filename checks. It also records that the validator differed from the approved design. This is a historical account; the current installation has not been tested for this page.
The guides describe review rules, note checks, and backups. The August 6 build notes record approvals into memory; current access restrictions and a recent restore have not been verified here.
CONNECTED WORK
More notes from the workbench.
Related write-ups on mistakes, checks, and local AI.