Agent OS · Component · personal-memory-management
Personal memory retrieval and lifecycle
What it does
Derives memory candidates from conversations/check-ins, deduplicates and ranks them, applies active windows/retention limits and supports manual create/edit/delete.
Core pattern
Memory extraction and deduplication + human edits and expiry
- Input
- Conversation/check-in text or manually approved memory
- Output
- Ranked context memories, write outcomes and memory summaries
Key parameters
| Name | Type | Default | Values |
|---|---|---|---|
| operation | string | recall | derive · create · update · delete · recall · cleanup |
| query | — | — | — |
| dedupe_threshold | number | 0.85 | 0–1 |
| limit | integer | 10 | 1–1000 |
| source_ref | — | — | — |
Maturity and limits
Repository upsert, expiry/pruning, candidate derivation and tests are implemented. Source paths were inspected statically; tests and live integrations were not executed. This rating describes implemented code, not production validation.
Highly personal data; retrieval uses rules/tags rather than general semantic embeddings; consent and retention need review.
Used in pipelines
Industry tags
- Lifestyle & consumer appsSource-supported industry
For work involving opt-in interactions, digital services and personal preferences, use memory extraction and deduplication followed by human edits and expiry.
- EducationTransferable use
For opt-in practice and role dialogue for adult learners, use memory extraction and deduplication followed by human edits and expiry.
Skills pack
Contract scaffold and deterministic mock
Download · 11KBSHA-256 2a521ae68cc8bdd2cc7bdd7f8ca300689d417639114ec749522d06d76b08ad4f
Related components
-
Versioned agent memory and knowledge lifecycle
Scoped memory retrieval + versioning and forgetting
Working 1 pipeline -
Streaming travel concierge
Conversational intake + tool search and candidate cards
Working 1 pipeline -
Contextual spiritual companion responses
Emotional cues and memory + supportive responses
Working 1 pipeline -
Onboarding and adaptive preferences
Explicit preferences + historically tracked preference signals
Working 1 pipeline