Skip to content

memory-capture.md

Latest commit

 

History

History
41 lines (34 loc) · 2.03 KB

memory-capture.md

File metadata and controls

41 lines (34 loc) · 2.03 KB
Raw

Save a lesson and recover it in another process

memory capture requires the actual lesson, either as --content or a UTF-8 --file. A provenance tag identifies the source; it is not a substitute for the lesson. Blank, unreadable and conflicting inputs fail without writing a memory. Content is preserved exactly, up to the evolution crate's 1 MB limit.

export TERRAPHIM_DEFAULT_DATA_PATH="$PWD/.terraphim-demo-data"
terraphim-agent --format json memory capture \
  --content 'Validate the garden brief with python3 validate.py --brief brief.md.' \
  --provenance-tag public-brief-v1
terraphim-agent --format json memory list
terraphim-agent --format json memory show MEMORY_ID_FROM_CAPTURE
terraphim-agent memory export --format json --output memories.json

These commands each start a new process. With the explicit data root, the evolution store is .terraphim-demo-data/terraphim/evolution/cli-agent.json. Without it, the historical platform configuration-directory path is retained; existing stores are not moved or imported. Use a different data root for a separate project store. This is file-location isolation, not an access-control system or automatic secret redaction.

To save a multiline note, use memory capture --file lesson.md. Text and file input are mutually exclusive. Earlier versions accepted capture with only a provenance tag and saved placeholder text; callers must now supply content.

Retrieval is a separate operation using the selected role's knowledge graph:

terraphim-agent --config role-config.json memory retrieve 'validate brief' \
  --role 'Brief Reviewer' --format json

Configure that role and its graph before retrieval. Saving a lesson does not create terminology or copy it into the separate learn store. Graph retrieval requires concept co-occurrence in the lesson; a role that does not recognise the query returns no memories rather than a lexical fallback. A role is a retrieval configuration, not a confidentiality boundary. memory apply is a preview of proposed context and does not execute commands.