Session end

Session End: Knowledge Base Sync + Memory Capture#

Captures curated learnings and mirrors them into a linked knowledge base (for example, an Obsidian vault organized Zettelkasten-style).

Adapt the paths and tool names below to your own setup. This skill assumes a daily-log knowledge base, a cross-session memory layer, and an optional memory plugin. The <vault>/ placeholder is wherever your knowledge base lives.

How this skill relates to a memory plugin#

A memory plugin (for example, Claude Mem) automatically generates a semantic session summary at session end and stores it in its own database. This skill does not duplicate that synthesis. Instead it captures curated signal: decisions that matter, memory updates, knowledge-base mirrors, and promotion candidates. The rich observation log stays in the memory plugin (query it with its search tool).

Step 1: Gather curated session context#

Don't narrate the whole session, the memory plugin has that. Extract:

• Key levers moved: 1-3 decisions or strategic shifts that matter
• Open loops: blockers, pending items carrying forward
• Next focus: priority for next session
• Projects touched: for knowledge-base links
• People mentioned: for knowledge-base links

Step 2: Daily log (primary record)#

Write or append to <vault>/daily/YYYY-MM-DD.md. Create the file if missing with frontmatter:

---
type: session-log
date: YYYY-MM-DD
---

Append a block for this session:

## [HH:MM], [workspace], [Brief title]

**Focus:** [one-line]

**Key levers:**
- [Decision or shift that matters]

**Open loops:**
- [Items carrying forward]

**Projects touched:** [[projects/slug-1]], [[projects/slug-2]]
**People:** [[name-1]], [[name-2]]

**Next focus:** [One line]

---

Keep it tight. If nothing moved, write one line: No significant levers, routine execution session.

Step 3: Thin local pointer#

Append a short pointer to your local session log at logs/sessions/YYYY-MM-DD.md so a weekly review tool can still find it without re-synthesizing what the memory plugin already stored:

## [HH:MM], [Brief title]

**Focus:** [one-line]

**Key levers:**
- [Top 1-3 items, verbatim copy from the daily log above]

**Next:** [Next focus]

_(Full session history in the memory plugin. Curated atoms live in the knowledge base.)_

---

Step 4: Update agent state#

Resolve the current user, then update the per-user state file (NOT the shared one):

{
  "last_updated": "YYYY-MM-DD",
  "current_focus": { "initiative": "...", "session_goal": "...", "notes": null },
  "last_session": { "date": "...", "focus": "...", "key_levers": [...], "open_items": [...], "next_steps": [...] },
  "last_session_note": "narrative paragraph capturing the session's thread"
}

Also update active_goals, reminders, and per-user recurring_tasks if they changed.

Never write personal session state to the shared state file, that file is the shared company source of truth (company block plus shared recurring tasks only). Each user owns their own session state in their per-user file.

Step 5: Memory files + knowledge-base mirror#

Review the session for information that should persist across conversations.

Mirror frontmatter:

---
type: project | reference
workspace: [workspace]
source_memory: <original memory filename, with .md>
aliases:
  - <original memory filename, without .md>
last_updated: YYYY-MM-DD
---

Body: verbatim copy of memory file content (drop original frontmatter).

Slug rule: filename without prefix, underscores → hyphens, no .md. For example project_ph_takeover.md → ph-takeover.md.

Why the alias (load-bearing): memory-file bodies link with the full slug, for example [[reference_cac_canonical]], but the mirror filename is hyphenated, for example cac-canonical.md. A linked knowledge base like Obsidian resolves [[slug]] against a note's aliases, so the alias line is what keeps those links from breaking. A mirror written without it lands as an orphan. Note: feedback_* files mirror to references/ with type: reference.

Connect it, don't just drop it. After writing a mirror, link it so it is not an orphan:

• Add it to the matching domain map-of-content (MOC) index.
• Make sure the daily-log block forward-links it under Projects touched or a References touched line.

A mirror with no inbound link is debt that a graph-health check will flag.

Memory file rules:

• The memory index stays small; one line per file, kept short
• Check for existing memory before creating new
• Update the index when files are added or removed

Step 6: Candidate detection#

Scan for promotion candidates. Append to your inbox candidates.md when any of these fire:

• Decision logged with explicit rationale
• New or updated feedback_* file
• Framework or mental model applied and worked
• User said "remember this," "important," or "worth noting"
• Memory flagged stale and the concept needs re-examination

Append under ## Pending:

## [YYYY-MM-DD workspace] {short title}
**Source:** memory/{filename} OR session log
**Snippet:** {1-3 lines, trimmed}
**Why surfaced:** decision | feedback | theme | explicit-flag
**Status:** pending

In-flow capture (the user's own words). When the session contains a passage where the user articulated a rule, framework, or tradeoff in their own words (chat or dictation), quote it back in the candidate entry verbatim, not paraphrased. Mark it so a later promotion pass can promote it without a fresh dictation pass:

## [YYYY-MM-DD workspace] {short title}
**Source:** session log (user's words, [HH:MM])
**Snippet:** "{verbatim quote, exact}"
**Why surfaced:** explicit-flag
**Status:** pending
**Promotion:** eligible for direct promotion (user's words)

Apply mechanical AI-tell strips only (no em dashes); never rephrase the quote.

Step 7: Surface promotion reminder#

Count Status: pending entries in candidates.md. If >5, end with:

N candidates waiting for a voice pass. Run the promotion skill when you want to turn them into permanent notes.

In-flow articulations. If this session produced any candidates marked **Promotion:** eligible for direct promotion (user's words), offer to act now rather than only queuing them.

Step 8: Permanent-instructions flag (optional)#

If a learning should become permanent instructions (not just memory), ask:

"Should any of these be added to the project instructions file for all future sessions?"

What NOT to do#

• Never re-synthesize the full session narrative. The memory plugin has the semantic summary. Extract key levers only.
• Never auto-write permanent voice notes. Those are the user's voice only, via a dedicated promotion pass.
• Never auto-edit the MOC index files. That is a separate synthesis step.
• Never rephrase memory file content when mirroring. Verbatim copy preserves voice.

Output format#

## Session Wrapped: [Date]

### Logged
- Daily log: `<vault>/daily/YYYY-MM-DD.md`
- Local pointer: `logs/sessions/YYYY-MM-DD.md`
- (Memory plugin semantic summary: automatic)

**Key levers:** [1-3 items]
**Open loops:** [Items]
**Next focus:** [Focus]

**Memory updated:**
- [Created/Updated/Removed: file, reason → mirror path]
- (or "No memory changes this session")

**Candidates surfaced:** N

[If pending > 5] **N candidates waiting for a voice pass.**

---
**Agent state updated.** Next session will start with this context.

Cross-references#

• session-start, next session; trusts the memory plugin injection for history
• A promotion skill, turn pending candidates into permanent voice-pass notes
• A synthesis skill, refresh the workspace index and MOCs