Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

Sign up
Appearance settings

waterdrop26651/Memento-skill

Repository files navigation

Memento Skill

简体中文

Skill skills.sh Validate Codex Claude Code No secrets License: MIT

Memory fragments banner

"I have to believe in a world outside my own mind."

"We all lie to ourselves to be happy."

Controlled recall for agents that forget, overfit, and hallucinate from old notes.

Memento Skill turns scattered experiment runs, notes, evidence, and hypotheses into a layered external memory system, so an agent can continue a project without treating every old fragment as current truth.

Don't let every note become a tattoo.

Copy This To Your Agent

Install https://github.com/waterdrop26651/Memento-skill and enable $memento-skill for future research-memory, experiment-tracking, and hypothesis-update work.

That is the preferred install path for most users: let the agent clone or install the skill into its own skills directory, then restart the agent if needed.

30-Second Demo

Before:

A7 looked better, A8 failed, maybe the retrieval change helped?
Old note says reranking was bad, but that might have been before the data fix.
Need next experiment.

After:

CURRENT_STATE.md -> what is actionable now
runs.csv -> A7/A8 facts, not interpretation
contrasts.csv -> prediction vs observed delta
hypotheses.md -> what would change the belief
ACTIVE_TRACKER.csv -> only evidence that still affects the next decision

Result: the next agent reads the hot path first, recalls archive only with a trigger, and proposes the highest-information contrast instead of replaying every note.

See examples/minimal-memory for a tiny before/after tracker.

What It Does

  • Keeps current evidence on a short hot path.
  • Separates facts, contrasts, and beliefs.
  • Archives stale branches without deleting them.
  • Recalls old memory only when there is a reason.
  • Ranks the next experiment by information gain.

Use It For

  • Long research projects with scattered runs and stale notes.
  • Agent handoff across sessions without full replay.
  • Hypothesis tracking where old evidence can poison new decisions.
  • Planning the next ablation, control, or contrast by information gain.

Memory Layout

CURRENT_STATE.md # current actionable reality
ACTIVE_TRACKER.csv # evidence with live decision gradient
EVIDENCE_LOG.md # compressed support for current beliefs
runs.csv # factual ledger
contrasts.csv # predictions, controls, outcomes
hypotheses.md # beliefs, risks, update rules
archive/ # cold memory
ARCHIVE_INDEX.md # recall map
RECALL_NOTES/ # recall audit trail

Install

CLI

npx skills add waterdrop26651/Memento-skill -g -a codex -y
npx skillsgate add waterdrop26651/Memento-skill

Manual Codex Install

mkdir -p ~/.codex/skills
git clone https://github.com/waterdrop26651/Memento-skill.git ~/.codex/skills/memento-skill

Then ask:

$memento-skill Build a controlled memory tracker for this project and rank the next highest-information contrasts.

Validate

python ~/.codex/skills/memento-skill/scripts/validate_tracker.py <tracker_dir>

Trust

  • No API keys or secrets required.
  • No background service.
  • Plain Markdown and CSV outputs.
  • Validation script included.
  • Designed for progressive disclosure: hot path first, archive only on trigger.

Skill Files

Share A Memory Layout

Using Memento Skill on a real project? Open a showcase issue with the hot path, the archived fragments, and the contrast it helped choose. Good examples make the skill sharper for everyone.

Banner image is an original generated asset, not a film still. Short quote sources: Memento (2000), see IMDb Quotes and Wikiquote.

AltStyle によって変換されたページ (->オリジナル) /