session-narrative

Session Narrative

You are writing a session narrative — a single flowing-prose memory note that captures the reasoning path of the current Claude Code session, not just the final decisions. Scattered memory_store calls capture what we decided; a session narrative captures how we got there — the attempts, the dead ends, the pivot moments, the lessons.

This skill is the Claude Code twin of the /session-narrative prompt file that ships with aman-copilot for VS Code Copilot Chat and Copilot CLI. Same protocol, same output shape, same memory store — works on all three surfaces transparently because the dev:* scopes share memory via amem.

When to refuse

If the session has been trivial — pure implementation of an already-decided plan, a 5-minute bug fix, or nothing surprising happened — tell the user the session isn't narrative-worthy and suggest scattered memory_store calls instead. Don't pad empty sessions into fake narratives. Narratives should preserve genuine reasoning; inflated narratives dilute the signal.

Good test: "Would a colleague joining the project next week actually benefit from reading this as a story?" If no, refuse and save facts instead.

How to write it

300–500 words of flowing prose. Not a bullet list. Prose, because prose preserves causation ("because X, we tried Y") in a way bullets can't. Write as if telling a colleague who joins the project next week and asks "how did we end up here?"

Cover in order:

1. Intent — what were we trying to do at the start?
2. Attempts — what did we try, in chronological order?
3. Dead ends — what didn't work, and why? This is the most valuable part.
4. Pivot moments — when and why did we change direction?
5. Outcome — what shipped, what's still open
6. Lessons — one or two reusable insights (or "none — this was execution")

Good narratives are honest about failure. If something took three attempts, say so. If a decision was a judgment call that might be wrong, flag it. The point is to preserve the thinking, not to write a success story.

How to save it

Option A — amem is installed (preferred)

Check if the amem-memory MCP server is available. If yes, call memory_store with:

• type: fact (amem doesn't have a dedicated narrative type yet — we use facts with clear title and metadata)
• confidence: 0.9
• content: the 300–500 word narrative, prepended with:

# Session narrative — <YYYY-MM-DD> — <short topic>

**Scope:** <current scope, e.g. dev:plugin>
**Type:** session_narrative
**Duration:** <rough estimate>

---

<the narrative body>

After storing, confirm with one line plus the narrative's opening sentence so the user can verify the right story was captured.

Option B — amem is not installed (fallback)

If amem is not installed, save the narrative to the Claude Code auto-memory directory at ~/.claude/projects/<current-project>/memory/session_<YYYY-MM-DD>_<short-topic>.md using the same markdown structure as Option A, with this frontmatter:

---
name: Session narrative — <YYYY-MM-DD> — <short topic>
description: <one-line summary of what this session was about>
type: session_narrative
---

Then tell the user: "amem isn't installed, so I wrote the narrative to your Claude Code auto-memory at <path>. If you install amem later, the next amem-cli sync will import this into your searchable memory store."

Safety

• Wrap secrets in <private>...</private> before any save — stripped before storage. This includes: API keys, tokens, URLs with embedded auth, and file paths containing credentials.
• Filter emotional context. If the user rejected a direction for reasons that were personal/tired/frustrated, keep the rejection in the narrative but frame it neutrally. Memory outlives the mood that produced it.
• Check consent for sensitive work. If the session touched client code, customer data, or production systems, ask before writing a narrative that might inadvertently describe those details.

If the user wants edits

If they push back — "you missed the part about X", "rewrite the lessons", "make it shorter" — revise in place and re-save via memory_patch (if using amem) or rewrite the file (if using auto-memory). Narratives should be singular per session, not duplicated. If they ask for a completely different framing, it's fine to delete the old one and write fresh.

What this enables

Three months from now, the user (or a future AI session) can run memory_recall("session narrative [topic]") and get back the entire story — not just the final decisions, but the reasoning path. That's the whole point. Decisions age poorly without their context; narratives age well because they carry their own reasoning.

This skill pairs with the amem prompt best practices guide at https://github.com/amanasmuei/amem/blob/main/docs/guides/prompt-best-practices.md which documents the full phrase catalog (save triggers, recall triggers, session closers) for users.