Session Wrap-Up

Agents start every conversation from zero. Re-pasting background costs tokens and loses detail. At the end of a session, write a short, structured record of what was decided and what is still open. At the start of the next session, read back only the entries that matter.

Credit: the pattern comes from a reel by nocodealex (2026-08-04, https://www.instagram.com/reel/Dbm200gvDsG/). It fits the episodic and semantic layers of the engineering workflow's memory rules.

Working context comes first

This skill is memory across sessions. Memory within a task is the working-context-notebook flow, which every multi-step task already runs. The two connect at both ends:

• At wrap-up, read the accepted notebook of each task in the session and take the decisions, verified results, and open tracker items from it. Do not rebuild them from the conversation. Copy only verified facts.
• At read-back, when the session continues a task, put what was recalled into a new notebook as unverified facts that name the session file as their source. Verify them against the current code before relying on them.

The notebook is disposable; the session file is the durable record.

Storage: one folder, one index

• Folder: <memory root>/sessions/. The default memory root is the device's global memory map when one exists (for example ~/Documents/Codex/global memory); otherwise ~/.agent-memory.
• One file per session, with the date in the name: YYYY-MM-DD-<agent>-<short-topic>.md. Never append to one growing file. A bad session can then be deleted without unpicking it from the others.
• One index: sessions/INDEX.md. Each entry is a single line: date, agent, project, topic, link, and 3–6 keywords.
• The folder is plain Markdown, so it can be synced to Google Drive and added to a NotebookLM notebook as a source. NotebookLM has no public write API: adding the folder is a one-time manual step (Drive sync or an upload), not something to automate or promise.

Wrap up (end of session)

1. Collect from this session only: decisions and their reasons, what was built or changed (paths, commits, links), verified results, failures and their fixes, open questions, and the next actions.
2. Write the file from the template below. Save decisions, not dialogue. No transcript, no pasted logs. Keep it under about 60 lines.
3. Record connections between tasks: which task's output feeds another, shared decisions, and contradictions with earlier sessions. These links are the most valuable part of the record, and they are why wrap-up happens per session rather than per task.
4. Add one line to INDEX.md.
5. If a lesson recurs (it is already in an earlier session file), promote it to a procedural rule (a skill, CLAUDE.md/AGENTS.md, or a hook) and say so.
6. Tell the user the file path in one line.

Read back (start of session)

1. Read INDEX.md only.
2. Pick the entries that match the current project, topic, or keywords (usually 1–3), and read those files. Do not load the whole folder.
3. Treat the content as leads: verify file paths, versions, and states against the current code before acting on them.

Never store

Secrets, tokens, passwords, API keys, client personal data (phones, emails, IDs), or account-owned transcripts. Record where a secret lives, never its value. Keep client names out unless the memory root is private to the user, and even then only as needed.

Template

# <YYYY-MM-DD> — <agent> — <project>: <topic>

## Decisions
- <decision> — because <reason>. (who decided)

## Done
- <change> — <path/commit/link> — verified by <test/command/screenshot>

## Failures and fixes
- <symptom> -> <verified cause> -> <fix>

## Connections
- <task A> feeds <task B> because <...>
- Contradicts/updates <earlier session file>: <what changed>

## Open
- <question or blocker> — waiting on <whom>

## Next
- <first action next session>

Index line format:

- 2026-10-08 · claude-code · ai-skills · engineering workflow + public repo · [file](2026-10-08-claude-code-workflow.md) · workflow, jury, memory, publish