plur-ai/plur/plur-session-end

PLUR Session End

Run this at the end of a conversation to extract what is worth remembering.

PLUR stores durable knowledge — corrections, conventions, preferences, decisions. It does not store session logs, one-off status, or anything that will be irrelevant next week. This skill enforces that discipline.

When to Use

• The user says "wrap up", "end session", "we're done", or similar
• The conversation is winding down after completing a task
• You want to record what you learned before the context window closes

Procedure

Step 1 — Identify learning candidates

Scan the conversation for:

Step 2 — Filter ruthlessly

Skip:

• Anything already obvious from reading the codebase
• Session-specific state ("we're working on the login page right now")
• One-off status ("the build was red this morning")
• Anything you are not confident about
• Secrets, credentials, API keys — never

Step 3 — Write each learning as a durable assertion

Format: a single clear statement that will make sense to a future agent with no session context.

Good: "The publish script requires npm 2FA — always run 'npm publish' interactively, never from CI."

Bad: "We talked about the publish issue."

Step 4 — Assign metadata

For each learning, determine:

• type: behavioral | terminological | procedural | architectural — choose the MCP schema category; correction/preference/gotcha can be tags
• domain: the project, library, or topic area (e.g., "plur", "typescript", "project:acme")
• scope: the scope of applicability (e.g., "global", "project:acme")
• tags: 2–5 descriptive tags

Step 5 — Save

plur_learn and plur_session_end are direct tools in both lean and full MCP profiles. For less-common operations, use plur_admin with { "action": "help" } to discover the current actions and schemas.

Save with plur_learn:

plur_learn({ statement, type, domain, scope, tags })

Call once per learning. Do not batch into one call — separate engrams decay and strengthen independently.

Alternatively, pass unsaved learnings as engram_suggestions to plur_session_end, following its live schema. Do not submit the same learnings through both paths.

Quality bar

Fewer, stronger engrams beat many weak ones.

• Prefer no engram over a vague one
• One sentence per engram — if you need two sentences, split it into two engrams
• If you're unsure whether something is reusable, skip it

What a good wrap-up looks like

Session learnings saved (3 engrams):
 
1. [correction] The API returns timestamps in Unix seconds, not milliseconds.
   domain: project:acme | tags: api, timestamps
 
2. [preference] User prefers TypeScript strict mode — always enable in tsconfig.
   domain: typescript | tags: tsconfig, preferences
 
3. [gotcha] The staging deploy requires a manual cache bust at /admin/cache.
   domain: project:acme | tags: deploy, staging, ops

Keep the user-facing summary short. Show what you saved; do not narrate your reasoning.

Integration with plur-memory

This skill pairs with plur-memory:

• plur-memory guides recall and learning; automatic injection depends on the installed editor adapter
• plur-session-end runs once — it extracts and saves what the session produced

Together they close the memory loop: inject at start, learn at end.