Privacy, ignore rules, and safety
Engram is private by default. Several layers keep private context from leaking into durable memory or across profile boundaries.
Approval gate
Writes require human approval. Agents propose candidates; humans approve, reject, edit, or archive. Direct terminal CLI uses A/B/C. AI-agent chat uses yes / audit / cancel.
Ignore rules
Ignore rules are privacy controls. They hide irrelevant or sensitive entries from routing. Configure them in .engramignore and workspace memory config so private paths and patterns never enter the index.
Hashes
Hashes are integrity checks. They run before content is printed and reveal unsafe edits that bypassed the normal write flow.
Profiles
Profiles isolate company, client, and personal memory so external APIs or company-provided agents do not leak context across projects. See Profiles and scope resolution.
Secrets and injection scanning
At save time Engram checks:
- schema validation
- secret scan
- prompt-injection patterns
- path safety
Limits to know
Default Engram search is deterministic lexical search. engram search --semantic adds deterministic local similarity, not embedding-backed semantic search. Graph vectors are local hashed word vectors, not semantic embeddings. Contradiction detection is advisory. Encryption config exists, but encrypted storage is not implemented yet.
These limits are intentional to state clearly. Engram should tell users what is real today and what is future work.