Agent Skills: Tree Ring Memory

Tree Ring Memory

Use Tree Ring Memory as a lifecycle-aware memory layer, not as a transcript dump.

Tree Ring Memory preserves meaningful agent learning like tree rings:

• fresh work stays detailed
• older learning compresses into stable rings
• important warnings remain visible as scars
• durable truths become heartwood
• speculative future work stays as seeds
• sensitive data is blocked, redacted, or kept out by default

When To Recall

Recall memory before:

• starting or resuming a project
• changing architecture, storage, security, privacy, or release behavior
• repeating a workflow where prior failures may matter
• responding to a user correction
• making a decision that depends on previous preferences or constraints
• editing files in a repo that has a Tree Ring Memory or AGENTS.md contract
• closing out meaningful work and deciding what should be remembered

Use narrow queries with project scope when possible. Prefer source-linked, high-confidence, non-superseded results.

When To Remember

Store a memory when the information is likely to help future work:

• the user states a durable preference
• the user corrects the agent
• a decision is made and should survive the current session
• an implementation lesson is validated by tests or production behavior
• a failed approach should not be repeated
• a security, privacy, release, or data-loss warning appears
• a useful project convention is discovered
• a future idea should be revisited later

Keep memory concise. Store the lesson, decision, or warning, not the full conversation.

Use tree-ring evidence instead of plain remember when the lesson comes from an evaluation, checkpoint, experiment, branch, incident, or reviewed run artifact.

Use source adapters when project artifacts already contain structured guidance or evaluated outcomes:

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

Run adapter commands with --dry-run first. Sync only concise, source-linked summaries; never treat imported memory as more authoritative than the source AGENTS.md, Revolve record, evaluation, PR, issue, or test artifact. In a Coordinated store, persisting an adapter result requires the coordinator capability; dry-run discovery does not.

Use the exact CLI commands exposed by the local install:

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

If the project was initialized with a project-local binary, prefer the generated .tree-ring/CLI.md reference and include --root .tree-ring when needed.

If this skill was loaded through a harness-native bridge file, treat that bridge as a pointer only. Read the project-local .tree-ring/SKILL.md and .tree-ring/CLI.md when present so commands match the installed project root. Do not assume a global Tree Ring setup applies to the current repo unless the user explicitly configured it.

Harness Activation

For a new project, begin with the safe, project-local default:

tree-ring init
tree-ring integrations status

Do not ask the user to copy a bridge or run integrations link for ordinary setup. init configures only safe project-local adapter material by creating absent final bridge and manifest paths. It never replaces or removes an existing entry, including during deactivation; contested entries stay untouched and report needs-user-review. A bridge, marker, generated skill, or successful init is not activation proof: active requires a fresh, matching receipt from a new session's scoped recall and safe context injection. Treat configured-awaiting-proof, active-isolated, needs-trust, needs-project-mount, needs-plugin, needs-user-review, unsupported, and failed as their exact non-active outcomes. Never say Hermes or another unverified runtime is active.

If publication durability becomes indeterminate, do not delete or rewrite the published path. Preserve disk material, keep changed harnesses marked needs-user-review in the returned in-memory manifest, and leave any activation manifest already published on disk intact for explicit reconciliation.

Pi trust is the user's decision: report needs-trust rather than changing global trust. Agent Zero is separate: tree-ring init writes only Tree Ring's passive Agent Zero binding with needs-plugin. The user installs/enables the compatible tree_ring_memory plugin and selects the mounted project; the plugin then owns its absolute, non-project activation-capability.json descriptor and passes it internally. Only descriptor-scoped plugin status can derive configured-awaiting-proof, and only its new-session preflight receipt can make the runtime active.

Never create a generic marker, copy or hand-author that descriptor, set its internal transport, modify Agent Zero core, or call a different plugin store shared. A missing, invalid, disabled, or release-incompatible descriptor stays needs-plugin; a different reachable store is active-isolated; an unavailable root is needs-project-mount. A passive binding, source checkout, or stale bundled CLI is not installed capability.

Receipts prove a privacy-safe preflight check, not durable memory creation or a security boundary. They exclude raw prompts, recalled content, secrets, sensitive values, paths, and coordinator capabilities. Shared-store claims are limited to same-host local-filesystem processes whose receipts match the canonical project store_id; they do not apply across hosts or network filesystems. For diagnostics use tree-ring integrations status --verbose; for advanced controlled work use integrations activate --harness <id> --dry-run, integrations certify, or integrations deactivate --harness <id>.

Evidence outcome mapping:

• promoted: durable heartwood from supported evidence
• rejected: scar for reusable failed or rolled-back approaches
• deferred: seed for promising unresolved options
• observed: outer-ring evaluation result

Memory Quality Gates

Use these gates before relying on or writing memory.

Recall gates:

• Before substantial project work, recall project constraints, scars, user preferences, and unresolved seeds.
• Before risky changes, recall warnings and evidence-linked prior failures.
• Before repeating a workflow, recall prior errors and accepted procedures.
• Before closeout, recall recent decisions so memory updates do not contradict already-stored lessons.

Trust gates:

• Prefer source-linked, non-superseded, high-confidence memories.
• Treat heartwood as durable only when source evidence or user confirmation supports it.
• Re-read source files, tests, explicit user instructions, DOX contracts, or Revolve evidence when memory conflicts with current sources.
• Do not treat sensitive or hidden-by-default memory as ordinary recall context.

Write gates:

• Remember only durable decisions, validated lessons, reusable warnings, corrections, future seeds, and evidence-backed outcomes.
• Reject transient planning chatter, duplicate wording, tool noise, and unsupported claims.
• Require evidence refs for promoted or rejected evaluated outcomes.
• Require user confirmation before creating or promoting broad cross-project heartwood.

Ring Selection

Use these rings:

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

Do not promote to heartwood from weak evidence. Prefer outer or seed unless the user confirms durability or the evidence is strong.

Event Types

Prefer specific event types:

• user_preference
• decision
• lesson
• warning
• correction
• file_change
• tool_result
• summary
• hypothesis

If a host integration has stricter event type names, use the closest local equivalent.

What Not To Store

Do not store:

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

If a useful memory contains sensitive material, store a redacted summary with enough context to be useful.

Source And Scope

Set project and scope deliberately:

• use project scope for repo-specific rules, decisions, warnings, and lessons
• use agent scope for agent-partitioned behavior and always set agent_profile
• use workflow scope for one coordinated fan-out/fan-in and always set workflow_id
• use session scope for one execution attempt and always set session_id
• use global scope only for durable user preferences or cross-project guidance
• include source references such as file paths, issue ids, PR ids, run ids, or docs paths
• use tree-ring evidence ... --evidence-ref <ref> for evaluated outcomes
• use tree-ring dox sync for concise AGENTS.md summaries
• use tree-ring revolve sync for promoted, rejected, deferred, or observed evaluation records
• use tree-ring integrations scan before configuring a new agent harness

Memory does not replace source documents. If a repo has AGENTS.md, project docs, tests, architectural records, or host-specific instruction files, read those sources directly and treat them as authoritative.

When DOX or Revolve source records change, re-run the matching sync adapter with --dry-run, inspect the generated memories, then run the write command only when the summaries are useful and source-linked.

Multi-Agent Coordination

For workers sharing one local Tree Ring root, give every write explicit coordination metadata:

tree-ring --root .tree-ring remember "Worker validated the storage boundary." \
  --event-type lesson \
  --scope agent \
  --project example-service \
  --agent-profile worker-storage \
  --workflow-id release-readiness \
  --session-id attempt-1 \
  --operation-id validate-storage-v1 \
  --source-ref runs/release-readiness/worker-storage.json

Use a unique agent_profile per worker, one shared workflow_id for the fan-out/fan-in, one session_id for each genuine execution attempt, and a stable unique operation_id for each logical write. An exact retry reuses both the original session ID and operation ID; changing only the session is a conflicting reuse. Start a new session and use new operation IDs only for a genuinely new attempt. Exact retries with the same operation metadata and payload return the original memory. Reusing that operation key for a different payload fails closed. Replacing a stored memory keeps its old operation namespace claimed. Redaction also tombstones the memory ID; only an explicit hard delete releases those claims.

At fan-in, recall the shared workflow and session without an agent-profile filter, inspect the source refs, then write a source-linked workflow or project summary:

tree-ring --root .tree-ring recall "release readiness" \
  --project example-service \
  --workflow-id release-readiness \
  --session-id attempt-1 \
  --scope agent

TREE_RING_AGENT_PROFILE, TREE_RING_WORKFLOW_ID, and TREE_RING_SESSION_ID provide the same defaults as their CLI flags. Do not leave an agent-profile environment filter set when the coordinator intends to recall every worker.

This shared-root pattern is for concurrent processes on one host using a local filesystem. It is not a distributed lock service and does not claim safe cross-host or NFS operation. Scope and identity fields remain routing metadata, not a read ACL; a same-user coordinator can recall across profiles. Use per-host stores plus an explicit, evidence-preserving fan-in process when work spans hosts.

Coordinated Write Policy

Stores default to backward-compatible Open mode. For a shared root where only a designated coordinator should publish or mutate shared memory, enable the optional Coordinated policy:

tree-ring --root .tree-ring policy enable --coordinator release-coordinator
export TREE_RING_COORDINATOR_TOKEN='<one-time capability printed by enable>'
tree-ring --root .tree-ring policy status
tree-ring --root .tree-ring policy audit --limit 100

Enable prints the capability once. Put it only in TREE_RING_COORDINATOR_TOKEN; never pass it as a CLI flag or place it in a memory, log, source ref, transcript, or committed file. Tree Ring stores only a hash. policy status and policy audit are read-only and do not reveal the capability. Inject it only into coordinator processes, and launch every ordinary worker with TREE_RING_COORDINATOR_TOKEN unset so fan-out does not inherit coordinator authority.

In Coordinated mode, an ordinary worker may only create non-heartwood scope=agent memory whose agent_profile matches its write context. Supply the same identity with --agent-profile <worker> or TREE_RING_AGENT_PROFILE=<worker>. A coordinator capability is required for:

• project, global, workflow, session, or other shared/non-agent writes
• heartwood creation or promotion
• JSONL import and persisted DOX/Revolve sync
• persisted consolidation
• ring changes and supersede/delete/redact lifecycle operations
• maintenance with apply or repair flags

Recall, export, policy status/audit, adapter dry-runs, consolidation dry-runs, and report-only maintenance remain read-only. In the TUI, start with --agent-profile <worker> (or TREE_RING_AGENT_PROFILE) so /remember defaults to agent scope. TUI promote/scar/seed, supersede, forget/redact, and persisted consolidation actions require TREE_RING_COORDINATOR_TOKEN.

Rotate the capability while the current one is exported, then immediately replace the environment value with the newly printed capability:

tree-ring --root .tree-ring policy rotate --coordinator release-coordinator-next
export TREE_RING_COORDINATOR_TOKEN='<new one-time capability>'
tree-ring --root .tree-ring policy disable
unset TREE_RING_COORDINATOR_TOKEN

Rotation invalidates the old capability. Disabling returns the store to Open mode and also requires the current capability.

This is operational write authorization enforced by official Rust/CLI store paths. It is not a read ACL, an operating-system security boundary, or protection from an adversary who controls the local database files or process environment.

Before opening an existing store with v0.13/schema v3, stop every Tree Ring process, checkpoint and back up the database, and upgrade every CLI, plugin, and bundled worker. Do not reopen the upgraded root with v0.12: schema v3 fences memory inserts, updates, and deletes from old writers, and all mixed-version operation is unsupported. Roll back only by stopping all processes and restoring the pre-upgrade backup.

Agent-Mediated Updates

Tree Ring Memory does not autonomously scrape chats or write durable memory in the background. The active agent is responsible for deciding when a Tree Ring command is warranted, then calling the CLI deliberately.

Use bridge files only to discover Tree Ring and its command reference:

• project-level bridges should point to .tree-ring/SKILL.md and .tree-ring/CLI.md
• global bridges should be treated as opt-in user configuration
• TUI event-stream pulses are display signals, not durable memories

Before writing memory, verify the lesson is durable, useful, privacy-safe, and grounded in user instruction or source evidence.

Forgetting And Correction

If memory is wrong, private, stale, or superseded:

• redact it when the durable shape is useful but details are unsafe
• delete it when it should not be retained
• supersede it when a newer decision replaces it
• prefer explicit reasons for every forget operation

In Coordinated mode these lifecycle writes require the coordinator capability.

Treat redaction as monotonic. Do not try to restore a redacted ID through replacement import; create a new reviewed memory only if the user deliberately reintroduces safe content.

Never keep known-wrong memory merely because it was previously recalled.

Closeout Habit

At the end of 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 sensitive and better left unstored?

Only remember the answers that will materially improve future work.