agentsclimarketplace
Sign in

Consolidate memory

Skill Myra-Agents/skills/skills/consolidate-memory

Extract persistent learnings from the current session and save them to the auto-memory system. Triggers when the user says "consolidate", "consolidate learnings", "save learnings", "what did we learn", "save what we learned", "extract memories from this session", or invokes /consolidate-memory. Also trigger proactively after a long debugging session, a corrected approach, or when the user explicitly validates something non-obvious. Do NOT wait for the user to ask if the session clearly produced memorable learnings.From its SKILL.md

Install
npx -y skills add Myra-Agents/skills --skill consolidate-memory

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.3 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it

Consolidate Memory

Scan the current conversation, extract what's worth keeping across future sessions, and write it to the project's auto-memory system.

read context → classify signals → draft memories → write files → update MEMORY.md

1. Compute the memory directory

MEMORY_DIR="$HOME/.claude/projects/$(pwd | sed 's|^/||; s|/|-|g')/memory"

Verify it exists: ls "$MEMORY_DIR". If the directory doesn't exist, the project has no memory system — tell the user and stop.

2. Check what's already documented

Before scanning for signals, grep the CLAUDE.md files to know what's already written down — anything captured there does NOT need a memory file:

# Find all CLAUDE.md files in the project tree
find . -name "CLAUDE.md" -not -path "*/node_modules/*" | head -10
# Then grep them for relevant terms as you identify candidate learnings
grep -l "git clone\|TUI\|bootstrap" CLAUDE.md */CLAUDE.md 2>/dev/null

If a learning is already stated in any CLAUDE.md, skip it. Memory files are for what code/docs can't capture — behavioral preferences, incident context, team decisions, non-obvious constraints.

3. Scan for signals worth keeping

Read the entire visible conversation. Extract only what a future Claude instance couldn't derive from the code, git history, or CLAUDE.md files.

High-value signals:

  • User corrections — "no, not that", "stop doing X", explicit pushback on an approach you chose
  • Validated non-obvious choices — user accepted an unusual approach without complaint (silence after an unexpected call = implicit confirmation)
  • "Aha" moments — a discovery that changed direction or resolved a persistent confusion
  • Project facts — deadlines, constraints, architectural decisions, external dependencies not obvious from code
  • User profile updates — revealed expertise, domain, workflow preferences

Low-value / skip:

  • Code patterns already visible in files
  • Git history (use git log instead)
  • Temporary state or in-progress work
  • Things already in MEMORY.md (check before writing)
  • Things already in any CLAUDE.md file (checked above)
  • Debugging recipes (the fix is in the code)

4. Classify each learning

SignalType
Correction / "don't do X" / behavioral preferencefeedback
Project fact, deadline, architectural decision, incidentproject
User role, expertise, workflow preferenceuser
Pointer to external resource, location of inforeference

5. Draft each memory file

Read the current MEMORY.md index to avoid duplicates. For an existing entry, update the file rather than creating a new one.

Filename: <type>_<slug>.md — short, lowercase, kebab-case. Examples: feedback_no_mock_db.md, project_release_freeze.md, user_go_expert.md.

File format:

---
name: <kebab-case-slug>
description: <one-line summary — specific enough to judge relevance cold>
metadata:
  type: <feedback|project|user|reference>
---

<body — see structure below>

Body structure by type:

feedback:

<The rule itself — what to do or avoid.>

**Why:** <reason the user gave, or inferred from context>
**How to apply:** <when this kicks in, scope, edge cases>

project:

<The fact or decision.>

**Why:** <motivation — constraint, deadline, stakeholder>
**How to apply:** <how this should shape future suggestions>

user: free-form, focused on role / expertise / preferences relevant to collaboration.

reference: <resource> at <location> — <what it's used for>.

Convert any relative dates to absolute (e.g. "Thursday" → "2026-06-12").

Link related memories with [[their-name]] slugs in the body.

6. Write files + update MEMORY.md

Write each new or updated file. Then update $MEMORY_DIR/MEMORY.md:

  • New entry: append - [Title](filename.md) — <one-line hook> under the relevant section (or at the end). Keep entries under ~150 chars.
  • Updated entry: edit the existing line if the hook is now stale.

Keep MEMORY.md under 200 lines total.

7. Report to user

List what you wrote, one line each:

Saved 3 memories:
• feedback_no_mock_db.md — don't mock the DB in integration tests
• project_auth_rewrite.md — auth middleware rewrite driven by compliance, not tech debt
• user_go_expert.md — deep Go expertise, new to React

If nothing was worth saving, say so plainly: "Nothing novel this session — all learnings already captured or derivable from code."

Edge cases

  • Nothing to save: don't create empty or trivial files.
  • Ambiguous signal: prefer feedback over project if in doubt.
  • Multiple learnings same topic: merge into one file rather than fragmenting.
  • Sensitive info (credentials, PII): never write to memory. Flag it to the user if found in the conversation.

What ships with it: 3 files

3.6 KB alongside SKILL.md

evals/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.