Tree Ring Memory

Overview

Use this skill to operate Tree Ring Memory as a lifecycle-aware memory layer for AI agent work. Tree Ring Memory is for durable decisions, lessons, warnings, project conventions, user preferences, and future seeds. It is not a transcript dump or a background scraper.

The core idea is that agent memory should age deliberately:

• fresh work can stay detailed while it is still active
• older lessons should compress into stable summaries
• important failures and warnings should remain visible
• durable preferences and project truths should become high-confidence memory
• speculative follow-ups should stay separate from confirmed facts
• sensitive data should be blocked, redacted, or forgotten

When to Use

Use this skill when:

• The user asks the agent to remember, recall, consolidate, redact, or forget.
• A task depends on previous project decisions, preferences, or warnings.
• The agent is starting or resuming work in a repository with Tree Ring Memory or a project-local .tree-ring directory.
• A test, incident, PR, benchmark, or review produces a lesson that should help future work.
• A source document such as AGENTS.md, DOX, or Revolve contains durable guidance that should be summarized into memory.
• The agent needs to audit stored memory before a risky change.

Do not use

Do not use this skill as the primary guide for:

• Short-lived scratch notes that should disappear after the task.
• Raw chain of thought or hidden reasoning.
• Secrets, credentials, tokens, private keys, payment details, or other sensitive values.
• Saving entire conversations instead of concise lessons or decisions.
• Treating unverified claims as durable project truth.
• Replacing source documents, tests, issues, PRs, or release records.

Instructions

Follow the workflow below whenever Tree Ring Memory could improve continuity. For small tasks, recall narrowly and only write memory when the lesson is clearly durable. For higher-risk work, include source checks, evidence-linked capture, and a closeout review.

Workflow

1. Recall before acting when prior context could affect the task.
2. Prefer narrow project-scoped queries over broad global recall.
3. Read source documents directly when they exist; memory does not replace AGENTS.md, project docs, tests, issues, PRs, or release records.
4. Store only concise lessons, decisions, warnings, and preferences that will materially improve future work.
5. Use evidence-linked capture when a lesson comes from a reviewed run, evaluation, checkpoint, incident, branch, PR, issue, or test artifact.
6. Redact, supersede, or delete stale or sensitive memory instead of preserving known-wrong context.

Command Reference

Start with local help so commands match the installed version:

tree-ring --help
tree-ring evidence --help
tree-ring dox sync --help
tree-ring revolve sync --help

If the project has a local Tree Ring setup, read .tree-ring/SKILL.md and .tree-ring/CLI.md before assuming a global configuration. If a command needs the project store explicitly, include the local root:

tree-ring --root .tree-ring recall --query "release decisions"
tree-ring --root .tree-ring evidence --help

Run source adapters in dry-run mode before writing imported summaries:

tree-ring dox sync --source-root . --dry-run
tree-ring revolve sync --source-root revolve --dry-run
tree-ring integrations scan --source-root .

Only write summaries that are concise, useful, source-linked, and privacy-safe.

Ring Model

Use the ring metaphor to decide retention strength:

• cambium: active task context
• outer: recent decisions and lessons
• inner: older compressed project knowledge
• heartwood: durable high-confidence truths and preferences
• scar: important failures, regressions, rejected approaches, and warnings
• seed: unresolved ideas, hypotheses, and follow-ups

Do not promote weak evidence into heartwood. Use outer or seed until the user confirms durability or the evidence is strong.

Privacy Guardrails

Never store:

• secrets, credentials, tokens, private keys, or payment details
• raw chain of thought
• temporary scratchpad notes
• unverified claims as durable truth
• sensitive health, financial, legal, or personal identifier details without explicit user instruction
• copyrighted source text beyond short allowed excerpts

When useful memory contains sensitive material, keep only a redacted operational summary with enough context to avoid repeating the same mistake.

Closeout Checklist

Before ending meaningful work, ask:

• What did we decide?
• What did we learn?
• What should future agents avoid repeating?
• Did the user state a durable preference?
• Is there a future seed worth revisiting?
• Is any memory wrong, stale, private, or better left unstored?

Only remember answers that are durable, useful, source-grounded, and safe.