| name | reflect-and-distill |
| description | Use when the user asks to reflect, summarise lessons, consolidate memory, recap a session, or take stock after a long working stretch. Pulls recent records via memory_recall, groups them into the four auto-memory categories (user, feedback, project, reference), and asks the user which insights to retain. Does not bypass or reconfigure the existing Reflector. |
Reflect and Distill
Drive a structured reflection pass over the agent's mem0-backed memory using only memory_recall and the LLM's judgement. The output is a categorised summary the user can review.
Step 1 — Gather raw material
Call memory_recall multiple times with different framings, capping each at top_k=10. Suggested queries:
recent decisions — what was decided lately, by whom, why.user preferences — how the user wants to work.failures / things that did not work — known mistakes to avoid.project state — current ongoing work.
If the user mentions a specific theme, add a query around it.
Step 2 — Categorise into four buckets
Read each retrieved record and decide which of the four auto-memory categories it belongs to (these mirror the categories used by Claude Code's auto memory system):
| Category | Holds |
|---|
| user | Who the user is, role, expertise, recurring goals. Not the user's behaviour in this session — that goes to feedback. |
| feedback | Corrections, approvals, working preferences. "Don't do X." "Yes, exactly like that." Always include the why if known. |
| project | Decisions and context not derivable from code: in-flight initiatives, deadlines, stakeholder asks. |
| reference | Pointers to external systems: dashboards, channels, ticket trackers, docs URLs. |
A record that doesn't fit any of the four is probably an ephemeral observation; drop it.
Step 3 — Present the distilled summary
Show the user the four categories with bullet points under each. Format:
## User
- <fact 1>
- <fact 2>
## Feedback
- <rule 1> (Why: <reason>)
- <rule 2>
## Project
- <fact 1>
## Reference
- <pointer 1>
Keep each bullet to one line. If a category is empty for this run, omit the heading rather than printing "(none)".
Step 4 — Ask which to retain
Ask the user explicitly: "Which of these should I keep, refine, or drop?" Don't assume — surfacing the choice is the value of this skill over a passive Reflector run.
Step 5 — Persist the keepers
For each retained insight, the agent's normal memory write path takes over: when the next agent.run() saves its outcome, mem0 picks up the conversation context. Do not invoke the Reflector directly — its threshold-gated cycle continues to run on its own schedule and shouldn't be double-fired.
If the host application provides an explicit "save to memory" tool, use that for items the user wants pinned now rather than next-run.
Anti-patterns
- Querying
memory_recall once with a generic string — you'll miss most of the relevant material. Three to five framings is typical.
- Inventing "reflection insights" that aren't grounded in retrieved records. Every bullet must trace to at least one record.
- Bypassing the Reflector — this skill is additive to the agent's existing reflection cycle, not a replacement. The Reflector still runs on its own threshold.
- Trying to write directly to memory_type=reflection — that's the Reflector's tag. Distilled-and-retained items use whichever memory_type the host application already uses for user-confirmed facts.
