Memory Recap

Generate prioritized actions

Focus the report on improvement decisions; the profile should explain why the recommendations fit this user. Usually select 3–5 distinct actions. This is a useful target size, not a quota. When evidence is limited, offer low-cost experiments explicitly marked “to be validated” rather than inventing recurring problems or benefits.

Each action should answer the following without becoming a lengthy form:

• Why change it? Which evidence or correction supports it? Is this an observed problem or an opportunity that still needs validation?
• What changes concretely? What will differ from the current approach next time? Who does it, when, and through which existing entry point? Provide a short usable instruction or minimal action.
• How might it help? Explain the rework or handoff gap it could reduce, mark the inference, and do not invent percentage savings.
• How will it be evaluated? Choose observable signals; collect a baseline first if none exists. State the additional cost and stopping condition without shifting the entire validation burden to the user.

Prefer improvements to agent execution. For example, turn “you value real validation” into “record the target environment and one real input in the existing task record, then have the agent report acceptance results using that input.” That specifies an action; “keep valuing validation” merely repeats a preference.

Rank actions by evidence strength, potential impact, and implementation cost, and identify which one to try first. Reuse existing specs, task records, test entry points, and receipts rather than defaulting to new meetings, approvals, templates, or files. A clear small task needs only a one-sentence action constraint.

For recurring corrections, consider the smallest appropriate durable home, but inspect existing coverage first. Not every problem needs a new Skill: project contracts, rules, tests, tool fixes, and memories have different scopes. Recommending that these assets be created or modified does not authorize carrying out those changes.

Do not turn one positive example into a universal admission requirement, such as “every Skill must have a deterministic script.” For existing multi-agent or knowledge-reuse workflows, first identify how to test their incremental value or address an actual gap rather than recommending adoption again in different words.

Check each recommendation before including it:

1. Could it be sent unchanged to any user? If so, add specific evidence and an action, or remove it.
2. Does it merely repeat something the user already does well? If so, identify the missing step.
3. Does it turn an agent responsibility into a new user process, or conflict with examples where direct execution was requested?
4. Does it claim unmeasured benefits or infer model quality from storage volume?

Analysis and verification

Analyze small inputs directly. Split large inputs according to context capacity and output headroom, preserving continuous line spans. Each batch should extract both findings and candidate improvements before synthesis and deduplication. Do not produce only profile summaries and expect the final pass to invent recommendations. Report actual input coverage; a model saying “read everything” is not proof that it understood every record.

Check separately:

1. Inputs: Snapshot digests match the combined original text; all selected content enters the analysis input, and deduplicated aliases remain traceable.
2. References: Sources and line numbers are valid and belong to the corresponding input; final citations trace back to intermediate evidence. Do not casually combine two evidence spans into a broader range.
3. Judgment: Open the key source passages and check the basis for conclusions and recommendations, duplicate events, and whether “recorded” has become “proven.” Mechanical citation validation cannot replace this step.

Revise only concrete gaps, retaining drafts and receipts. Do not rerun the full corpus for local wording changes or retry indefinitely. If a new recommendation requires current code, runtime, or retrieval evidence to hold, leave it pending validation rather than guessing.

Delivery

Start with a short work profile, followed by prioritized actions and the single change most worth trying first. An optional display card must not replace the requested recommendations. Keep the report readable; put evidence tables, original text, and execution details in supporting files.

Deliver the recap, source manifest, and any requested original-text compilation. For external model execution, retain the actual model, completion status, invocation count, and usage. Distinguish estimates, receipts, and unknowns; do not interpret total_cost_usd: 0 as the absence of other billing units.

Explain the limits of what the materials can answer without letting lengthy disclaimers crowd out actions. Analysis itself does not authorize writing back, merging, or deleting native memories, or publishing private corpora.