# Consolidate Memory

> Clean up memory (merge duplicates, fix or remove stale entries) when the user asks.

- Skill: `invergent-ai/consolidate-memory` (Agent Skill)
- Install (CLI): `npx get-skill add invergent-ai/consolidate-memory`
- Raw SKILL.md: https://api.skillmd.com/api/skills/invergent-ai/consolidate-memory/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: invergent-ai (https://skillmd.com/u/invergent-ai)
- Updated: 2026-11-17
- Page: https://skillmd.com/skills/invergent-ai/consolidate-memory

---


# Memory Consolidation

You're doing a reflective pass over what you've learned about this user and their work. The goal: a future session should be able to orient quickly — who they work with, what they're focused on, how they like things done — without re-asking.

Memory is two lists of entries, edited with the `memory` tool: `user` (who the user is, their preferences) and `memory` (your notes on their projects, conventions and workflows). Both are injected into every turn of every future session, so every entry costs context forever.

## Phase 1 — Take stock

- Read both lists as your system prompt shows them. That is the state when this session started; account for anything you wrote since.
- Note which entries overlap, which look stale, which are thin, and which sit in the wrong target.

## Phase 2 — Consolidate

**Separate the durable from the dated.** Preferences, working style, key relationships, and recurring workflows are durable — keep and sharpen them. Specific projects, deadlines, and one-off tasks are dated — if the date has passed or the work is done, remove the entry or fold the lasting takeaway (e.g. "user prefers X format for launch docs") into a durable one.

**Merge overlaps.** If two entries describe the same person, project, or preference, `replace` the richer one with the merged text and `remove` the other.

**Fix time references.** Convert "next week", "this quarter", "by Friday" to absolute dates so they stay readable later.

**Drop what's easy to re-find.** Task progress, session outcomes, and completed-work logs don't belong in memory — `session_search` recalls those from past transcripts. Cut entries that restate something you could pull from the user's files or connected tools on demand. Keep what's hard to re-derive: stated preferences, context behind a decision, who to go to for what.

**Keep the user's word.** Never drop or soften a preference the user stated unless a newer statement of theirs contradicts it; when two contradict, keep the newer and say so in the summary.

**One fact per entry, written compactly.** `replace` identifies an entry by a short unique substring (`old_text`) — pick one that appears in that entry only.

Finish with a short summary: how many entries you merged, rewrote, and removed, and what changed.

