> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wazoo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory appraisal

> How to decide what belongs in agent memory, source systems, or nowhere.

Memory is not saved chat history. Memory is the result of deciding what has
durable value, where it belongs, and what should disappear after the durable
record exists.

Worlds treats agent memory as curated world state. A world should preserve the
facts, relationships, provenance, and preferences an agent needs to act later
without rereading every prior transcript.

## Appraise before storing

Before adding material to a world, decide what kind of record it is.

| Material             | Better destination                    | Why                                                                             |
| -------------------- | ------------------------------------- | ------------------------------------------------------------------------------- |
| Raw chat transcript  | Temporary logs or an external archive | Most turns are context for the moment, not durable state.                       |
| Stable preference    | World fact                            | Future agents should not ask again or infer it from stale chat.                 |
| Durable decision     | World fact with provenance            | Later actions need to know what was decided and why.                            |
| Source-of-truth data | The source system                     | Worlds should reference or synchronize the durable fact, not replace the owner. |
| Recurring blocker    | World fact or issue tracker           | Repeated friction should become inspectable state or actionable work.           |
| Duplicate summary    | Existing canonical record             | Extra summaries increase context burden.                                        |
| Temporary report     | Scratch space                         | Promote the useful conclusion, then discard the report.                         |

## Keep memory small enough to use

A useful world reduces future context burden. It should make later work cheaper
by preserving stable facts and deleting noise.

Good memory captures:

* durable facts and relationships
* stable preferences
* accepted decisions
* recurring blockers
* corrected assumptions
* provenance for claims that matter

Poor memory captures:

* full transcripts by default
* duplicate summaries
* temporary status reports
* stale hypotheses without evidence
* data that already belongs in another source system

## Preserve provenance

When a fact matters, store enough provenance to explain why it exists.
Provenance can point to a source document, an issue, an API record, an import
batch, or a human decision.

This lets agents answer different questions without guessing:

* What do we believe is true?
* Where did that fact come from?
* Which source owns the durable record?
* What changed since the last run?
* Which facts are stale or superseded?

## Use Worlds as the curated layer

Worlds is a substrate for curated state, not a dumping ground for every token an
agent has seen. Use imports, updates, search, and SPARQL to keep important facts
inspectable and queryable.

When another system is the source of truth, keep that boundary. Store the fact,
pointer, or synchronization result that helps the agent act, and leave the owned
record where it belongs.

## Next steps

* Use [Search](/worlds/search) to retrieve relevant facts from a world.
* Use [Query](/worlds/query) to inspect relationships and provenance.
* Use [Update](/worlds/update) to change durable state deliberately.