Design session-end memory suggestions and focused recall #77

Concept v0: session closeout

The active coding agent proposes at most a few durable findings after meaningful work: decision, gotcha or reusable note. Each proposal includes source evidence, project/scope, rationale, proposed summary/detail and an existing memory/revision if it updates prior knowledge.

Proposed flow

1. Search authorized existing memories and classify as new, update, duplicate or conflict.
2. Present a concise reviewable diff with source links; discard transient logs and secrets.
3. Save only according to the user's existing authority and moderation policy. A suggestion does not grant publish/review permission.
4. Preserve attribution and concurrency; avoid a second paid model call by default when the active agent can prepare the proposal.

Recall idea

Evaluate query relevance, current validity, project/scope and a strict token budget against the current baseline. Prefer linked decisions for the selected task; do not assert that embeddings are necessary or that important knowledge may be silently discarded.

Deliverable

A reviewable design and small evaluation set, not a production feature or new automatic data collection. Define explicit approval points, costs, deduplication and conflict behavior.

Acceptance

• Concrete new/update/duplicate/conflict examples and proposed request/response contracts.
• Privacy, revision and permission behavior specified.
• A measurable comparison of recall relevance and context size before selecting an algorithm.

Reference

Mem0 inspiration: https://github.com/mem0ai/mem0/blob/main/docs/core-concepts/how-it-works.mdx

Agent handoff

Design task only. Product implementation needs a later explicit task.

Concrete subtasks

1. #97 Session-closeout proposal contract and examples.
2. #98 Recall evaluation design.

Both are concept work and can proceed independently. This epic is complete when the designs are reviewable; it does not require shipping an automated memory feature.