Add, search and correct memories
Write text into your product's memory, ask it a question in ordinary language, page through what it holds, and fix or remove a statement that is wrong.
Everything on this page acts on one product’s memory. Use it to put a statement into memory, to get an answer out of it, and to repair what is wrong once it is in. What a memory is explains the model behind these operations — scope, source authority, and why one write produces several memories.
Agents in your product can do all of this, including writing; people read, correct and clean up, and do not create memories themselves. Each task below names the capability. The calls behind it are published by the platform: look up a Brain tool.
Add a memory
A memory is created by writing text into it. The writer supplies the text and chooses two things about it.
- Write the text. Everything else about the memory is derived from it.
- Choose the scope — shared with the whole product, or private to the writing agent. Shared is the default, and structured metadata can be attached to the write and is kept with the memory.
- Choose the source authority — how much weight the statement carries against a competing one.
officialis the default.
What comes back is the identifier of what was stored, which is what you use later to open, correct or delete it.
One write usually produces several memories: the text as submitted, plus the individual statements Brain read out of it. Memories also arrive without anyone writing them, from the sources you connect and from Brain’s own consolidation.
Search
Search takes a question in ordinary language and returns the memories that answer it, best match first. It is not keyword matching: a memory phrased differently from your query still comes back. You can set a result limit, and restrict a search to shared memories or to the searching agent’s private ones; by default it covers both.
Results come back in three groups, and every result carries its match score, its kind and its source authority.
| Group | What it holds |
|---|---|
| Profile | Standing background about the product’s world — who and what it deals with, and the patterns Brain has induced about it. True regardless of what you asked; read it as context for the rest. |
| Results | The direct answer to the query. This is where most of what you are looking for appears. |
| Proactive | Standing intentions Brain derived — things it holds as worth doing or watching. Relevant to the query, but pointing forward rather than answering. |
Any of the three can be empty. A product whose memory has not yet been consolidated typically has results only. See how memory maintains itself.
List memories
Listing walks the whole memory in order rather than answering a question. It pages — you choose the page size and the offset — and can be ordered oldest or newest first. It reports the true total for the scope you are reading, not just the size of the page in front of you, and like search it covers shared and private together unless you restrict it.
Only current memories are listed. A memory superseded by a newer statement does not appear; the memory that replaced it does.
Open one memory
Opening a memory by its identifier gives you the full record: the content, whether it is the current version, what it supersedes and what superseded it, its source authority, its kind, its tags and metadata, its confidence, and who wrote it.
You can only open a memory that belongs to your product, and only if it is shared or your own private one. Anything else reads as not found.
Correct a memory
Correcting replaces a memory’s text. Nothing else about it changes — not its scope, not its source authority, not its history, not when it was recorded. The correction is what is searched from then on.
| You want to | Do this |
|---|---|
| Fix a statement that is wrong or badly phrased | Correct the memory. |
| Record that the world has changed | Write a new memory. It supersedes the old one on its own. |
Delete a memory
Deleting removes the memory, and cannot be undone: recovery means restoring from a backup. Deleting repairs the supersession history around what it removes, so the chain leading to the current version does not break. As with opening, you can only delete a memory in your own product that is shared or your own private one.
Deleting a memory that Brain derived from a longer piece of text does not remove the others derived alongside it, and does not remove the original text. If you want the whole write gone, all of it has to go.
See the statistics
Brain reports summary counts for the product: how many memories it holds in total, how many of those are plain facts, summaries and identity statements, and how large the knowledge graph is — its number of nodes, its number of relations between them, and how many distinct entities it knows about. The counts follow the scope you read them for.
They are useful as a coarse health check. A product whose memory is being written to but whose entity count stays flat is worth looking into.
Next steps
- What a memory is: scope, source authority, and what a write actually stores.
- Search the knowledge graph: query the entities and relations the counts above report on.
- How memory maintains itself: why the profile and proactive groups fill in over time.
- Back up and restore memory: the only way back from a deletion.
- What Brain meters and which limits apply: what these operations consume, and the quotas on them.