Understand memory

Layered recall

Understand how memory search expands from a quick hit to deeper context.

Layered recall is how Agentic OS reads memory.

It starts with search. It only asks for more context when the first result is too small to answer safely.

A user question passes through a scope check, search combines semantic and keyword matches, expand reads nearby chunks, and local transcript drill-down reads exact conversation turns.
Layered memory recall

How it works

Recall has three layers:

LayerWhen to use itWhat it returns
SearchFirst step for every question.Ranked memory chunks from the allowed scope.
ExpandWhen the best result is too short.Nearby chunks from the same source.
TranscriptWhen exact wording, a command, or output matters.A small local transcript window around the matched turn.

Search combines semantic search and keyword search. Semantic search helps with similar meaning. Keyword search helps with exact words, IDs, filenames, and settings. The results are combined and ranked before recall returns them.

Expand uses the chunk_id from a search result. It reads nearby chunks from the same source and the same scope.

Transcript drill-down also uses a chunk_id. It checks the same scope again, then reads a small window from context/transcripts/. This is local-only for now because raw transcripts stay on the local machine.

Commands

Run search from command-centre/:

npm run memory:recall -- "onboarding notes"

Expand a search result:

npm run memory:recall -- --expand <chunk-id>

Read the local transcript window for a result:

npm run memory:recall -- --transcript <chunk-id> --local

Use the same scope flags across the whole recall. For example, if search used --client acme, use --client acme again for expand.

Hosted Team OS behavior

When the local workspace is signed in to Team OS, search and expand use the hosted API by default.

Hosted Team OS supports:

EndpointPurpose
POST /v1/memory/searchSearch allowed memory.
POST /v1/memory/expandExpand a search result into nearby source context.

Transcript drill-down still needs --local because it reads raw transcript files from the local workspace.

For API details, see Hosted Team OS API.

Next: Hosted Team OS API

On this page