agentleFS
mikeyobrienmikeyobrien/rho/skills/memory-consolidate/SKILL.md

memory-consolidate

Skill372 starsChanged 8 months agoMITView on GitHub

Consolidate brain memory and mine user sessions since the last consolidation checkpoint (sleep-cycle style). Use to reduce noisy prompt injection while preserving durable high-value memories.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 3 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run a "brain sleep cycle":

    line 11
  2. Very specific:

    Use the brain tool for brain changes and the vault tool for reference relocation.

    line 16
  3. Distinctive:

    Use these symbols while triaging entries:

    line 22
  4. Very specific:

    If since is explicit timestamp/duration, use it.

    line 57
  5. Very specific:

    If since=last_consolidation, read checkpoint from checkpoint_key.

    line 58
  6. Very specific:

    If no checkpoint exists, use days_fallback.

    line 59
  7. Distinctive:

    Leave a short pointer memory when useful (e.g., "See [[note-slug]]").

    line 127

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Memory Consolidate

Overview

Run a "brain sleep cycle":

  1. Consolidate existing memory (dedupe, decay, merge, vault relocation)
  2. Mine user sessions since the last consolidation checkpoint
  3. Persist a new checkpoint only after a successful run

Use the brain tool for brain changes and the vault tool for reference relocation. Never edit brain.jsonl directly.


Importance levels (retention)

Use these symbols while triaging entries:

  • 🔴 High importance — durable, high-leverage, should remain in brain
  • 🟡 Medium importance — useful but optional; review for merge/tightening
  • 🟢 Low importance — stale/noisy/duplicative; prune or relocate

For memory retention, high importance means keep longer, not delete.


Parameters

  • brain_path (default: ~/.rho/brain/brain.jsonl)
  • mine_sessions (default: true)
  • since (default: last_consolidation) — last_consolidation | <ISO timestamp> | <duration>
  • days_fallback (default: 1) — only used when no checkpoint exists
  • session_dir (default: ~/.pi/agent/sessions/)
  • max_new_entries (default: 10)
  • confidence_threshold (default: high) — high | medium
  • checkpoint_key (default: memory_consolidate.last_consolidated_at)

Steps

1) Inventory

List memory by type and count totals.

You MUST report counts for:

  • learnings, preferences, behaviors, identity, user, context, tasks, reminders
  • total active entries

2) Resolve mining window

Determine the lower bound timestamp:

  1. If since is explicit timestamp/duration, use it.
  2. If since=last_consolidation, read checkpoint from checkpoint_key.
  3. If no checkpoint exists, use days_fallback.

Constraints:

  • You MUST mine only sessions in the resolved window.
  • You MUST include the resolved window in the final report.

3) Session mining (user messages only)

Extract durable learnings/preferences from user messages in matched sessions.

Confidence policy:

  • High: explicit user statements/corrections/preferences → auto-add
  • Medium: strong multi-session inference → add only if threshold is medium
  • Low: ambiguous/one-off/hypothetical → skip

Constraints:

  • You MUST NOT exceed max_new_entries.
  • You MUST dedupe against existing memory before add.
  • You MUST prefer high-confidence extractions first.
  • You MUST include session id in source when available (session:<id>).

4) Decay stale learnings

Run brain action=decay.

Constraints:

  • You MUST report decayed count.

5) Consolidate existing entries

Identify duplicates/superseded/stale entries and merge candidates.

Apply the importance lens:

  • 🔴 Keep durable, high-value operational guidance
  • 🟡 Merge or tighten wording
  • 🟢 Remove if stale/noisy/redundant

Apply the 30-day test to every entry: "Would this change a decision I make 30 days from now?" If no, it's noise — remove it.

Auto-remove categories (these should never have been stored):

  • Version numbers or update confirmations ("updated X to v1.2.3")
  • Heartbeat or check-in status reports ("Heartbeat Feb 19: all clear")
  • Benchmark scores or run results ("scored 42/89 = 47.2%")
  • Bug sweep summaries without a generalizable root cause ("reviewed X, no bugs found")
  • UI/feature implementation details ("button text changed to X", "layout uses 3 columns")
  • Task completion status ("task X is complete", "run Y failed")
  • Project-specific transient state that won't inform future decisions
  • Duplicates — keep the best-worded version, remove the rest

Constraints:

  • You MUST NOT remove preferences unless contradicted/superseded.
  • You MUST NOT invent new facts while merging.
  • When uncertain, keep.
  • You SHOULD be aggressive about pruning — a smaller, high-signal brain is better than a large, noisy one.

6) Vault relocation for reference-heavy entries

"Reference-heavy" means useful knowledge that does not need to be injected every turn and can be searched ad hoc.

Typical candidates:

  • long feature histories / changelog-style learnings
  • architecture rationale requiring structure
  • multi-step runbooks / deep troubleshooting notes
  • linked research/reference material

Constraints:

  • You MUST write vault notes before removing corresponding brain entries.
  • Each note MUST include ## Connections with [[wikilinks]].
  • Leave a short pointer memory when useful (e.g., "See [[note-slug]]").

7) Persist checkpoint (success only)

At end of successful consolidation, set/update checkpoint timestamp (now, UTC ISO-8601).

Constraints:

  • You MUST update checkpoint only after successful completion.
  • You MUST NOT advance checkpoint on partial/failed runs.

8) Report

You MUST report:

  • counts before/after by type + total
  • mining window and sessions analyzed
  • added/skipped mined entries (with skip reasons)
  • decayed, removed, merged, relocated counts
  • vault notes created/updated (slugs)
  • checkpoint old → new value
  • up to 10 significant changes

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.