Governed memory for continuity across long-running work.
- Status
- Planned
- Priority
- 6 of 6
- Best use
- Long-running cases, prior decisions and user preferences, kept apart from source facts
- Deployment
- Standalone memory tool; use customer memory services when already present
- Relative effort
- M-L
Pick it when people come back to the same case over time
Best for
- Long-running cases and workflows that span days or weeks
- Remembering a user's preferences and earlier decisions
- Hand-offs: what was decided, by whom, and why
Ask it things like
- What did we decide about the Rogers credit dispute last week?
- Which report format does this user prefer?
- What are the open tasks on case 4471?
Not the right choice for
- Facts about the business (memory is never a source of truth; figures are reconciled with the sources)
- One-off questions with no history
Choose it over Agentic when the same people come back to the same case and need continuity. Memory adds what was decided; Agentic finds what the data says.
- Data it reads
- Memories a person approved before saving, scoped to a user or a case.
- How it gets the data
- Recalls memories by scope, keeps them in a separate section from the source evidence, and checks remembered figures against the sources.
- How fresh the answer is
- Whatever was approved last; old memories are marked outdated, not silently reused.
- Relative cost
- Low
How Memory RAG works
- 1
Select Candidate Memories
Identify eligible interaction facts, decisions, outcomes and preferences; avoid persisting entire conversations by default.
- 2
Classify & Govern
Attach purpose, scope, tenant/user IDs, provenance, time, sensitivity, retention, correction and deletion policy before writing.
- 3
Consolidate & Store
Deduplicate, resolve conflicts conservatively, preserve history, and store only allowed memory objects in a customer-approved store.
- 4
Recall by Scope
Retrieve relevant, authorized memories using purpose, entity, freshness and tenant/user bounds.
- 5
Reconcile with Authority
Check age and conflict against trusted systems when a memory asserts source-owned facts; mark uncertain or superseded content.
- 6
Draft with Distinctions
Use approved memory as context, clearly separating recollection, recorded decisions and currently authoritative facts.
- 7
Validate, Return & Maintain
Ground material claims where possible, enforce output policy, and update or expire memories only when governance allows.
Support opt-out, deletion, correction, TTL, tenant isolation and purpose limitation as configured by the customer.
One router picks the right method for each question
Context66 combines these methods as Ontology-guided Adaptive RAG. Every answer is checked against its evidence, and the system abstains when the evidence is incomplete.
- User question
- Intent and context detection
- Ontology and semantic interpretation
- Query router
All branches, orchestrated by the agent
- Evidence fusion and ranking
- Answer written only from retrieved evidence
- Claim-level validation and citation checking

