Design memory around the decision loop.
Support agents, coding agents, assistants, and companions all need continuity—but they should not share one undifferentiated storage recipe. Scope, write policy, source retention, and correction rules change with the loop.
- 01ObserveIdentify the outcome, correction, preference, or constraint that repeats.
- 02RememberStore a scoped conclusion and attach source provenance.
- 03ApplyRecall before the next relevant decision and record feedback.
- 01support continuity
- 02project decisions
- 03changing preferences
- 04source-backed answers
Start from the repeated decision the agent must make, then define the smallest durable record and evidence needed to make it better next time.
Resolve the next ticket with prior outcomes.
Remember verified customer preferences, troubleshooting steps that already failed, and the final resolution. Keep policies and manuals in Documents so the answer can cite its source.
- Scope customer facts by authenticated user or account.
- Invalidate outdated plan or device state instead of overwriting history invisibly.
- Attach ticket identifiers as provenance without storing unnecessary transcript text.
Carry architecture across sessions.
Persist decisions, conventions, failed approaches, and release constraints. The Desktop integration keeps this loop local; Cloud can share the same contract across a team service.
- Recall before planning or touching a subsystem with prior decisions.
- Store why a choice was made and where the source can be inspected.
- Correct stale instructions explicitly when the repository changes.
Track what is true now—and what changed.
Use bi-temporal records for location, preferences, relationships, and ongoing commitments. Separate sensitive user facts from general documents and derive scope from the signed-in user.
- Current recall can favor active facts while history remains inspectable.
- Feedback gives applications a correction signal without mutating results silently.
- Retention and deletion policy should be part of product UX, not only an API endpoint.
What this surface includes
- Runnable cookbooks for personal assistants, coding agents, companions, support agents, and document RAG.
- Patterns for scope, provenance, correction, events, and retrieval evidence.
- Deployment choices for one local user, a self-hosted service, or managed multi-tenant Cloud.
Explicit boundary
- Use-case guidance is an implementation pattern, not a claim about customer adoption or outcomes.
- Applications remain responsible for consent, retention, and sensitive-data policy.
- A memory write policy must be evaluated with the downstream task, not by record count.