# Adopt durable team records

Practices: typed-records, atomic-claims, durable-handoffs, decision-records.

Read [Team Memory and governed operational memory](https://hamedtaheri.com/playbook/v0.2.0/reference/operational-memory.md) for the purpose of these records. Organize continuity around the relevant entities, events, decisions and ongoing work, not only a conversation or one user's preferences. Session, personal and operational memory can overlap; lifetime, scope and purpose are separate dimensions. Preserve source attribution, time/validity where available, authorized scope and the authority behind decisions. Retrieve only what the current participant needs. Shared memory does not grant permission to act or replace the source system. Distinguish implemented controls from instructions and future capabilities.

Inventory the existing backend and schemas before adoption. Use the [kit definitions](https://hamedtaheri.com/playbook/v0.2.0/kit/definitions.json) to assess compatibility. Preserve local identity rules, ownership, statuses and custom fields. Add missing compatible definitions only when setup is authorized.

Read the [full team-kit architecture](https://hamedtaheri.com/playbook/v0.2.0/reference/kit-architecture.md) and [JSON schema map](https://hamedtaheri.com/playbook/v0.2.0/reference/kit-architecture.json) when assessing collections, entity types, document categories, relations or guidance. It is a source snapshot of AI Coding Memory 1.2.0, not the portable starter schema or a live installation receipt; the [kit map](https://hamedtaheri.com/playbook/v0.2.0/reference/kit-map.html) shows the current version (1.4.0 adds routine records; 1.5.0 makes the routine playbooks neutral templates to copy and adapt). Do not infer an edge from a field or a diagram: explicitly write and verify approved relationships. Use structured, fully paginated filters for complete work lists. Supporting specs/plans/handoffs are document content, not additional entity types in this kit.

Records need stable typed keys, versioned writes, scope-restricted retrieval and evidence references. Read before changing status/owner; a conflict means re-read and resolve, not retry with stale assumptions. Verify writes and say unsaved when they fail.

The local adapter demonstrates create, get, guarded update and task claim. It has no semantic recall or production permission model. Use it only for the documented private disposable exercise. The [provider mapping](https://hamedtaheri.com/playbook/v0.2.0/kit/personize-mapping.md) describes additional qualification for hosted storage.

At a handoff record goal, work references, revision, attempts, verified acceptance, unresolved questions and next step. Record human decisions with their authority. Suggestions are not tasks until the human accepts them.

Keep open questions and answers as separate records. A question only the human can answer is an inquiry directed to them, and the task waiting on it is blocked by it; a proposal stays an inquiry until accepted. An answer that a later session would otherwise re-ask or undo is a decision linked to the inquiry it resolves and the tasks it applies to. A one-time approval is a note on the task; a tactical choice inside one task needs no record. A changed answer is a new decision that supersedes the old one; never edit an accepted decision.

Word a decision as a findable general rule: a one-line title with domain terms, the general question, what is allowed with its limits, the rationale with the human's own words quoted briefly, the alternatives and the affected areas. Set a reuse level: `standing_rule` only when the human said "always" or "from now on" (give it a review date), `precedent` for a strong default, `one_off` otherwise. At pickup, list open inquiries, applicable accepted decisions and ranked precedents. Treat accepted decisions as final. Before opening a new inquiry, search for a precedent and offer it as the recommendation. Headless workers never ask: leave the inquiry open and hold the task. When importing older answers, tier them; do not turn every approval into a decision. Keep transcripts, secrets and customer data out of the public reference and contribution export.