How it works
How agentic memory becomes usable context.
Authorized signals become canonical identity, qualified memory and a task-bounded context card. Evidence and scope stay inspectable throughout.
Current mechanics are kept distinct from beta correction outcomes and conceptual cross-scope sharing.
Architecture principle
One published memory surface. Private state behind it.
Agents search safe published memory. Source generations, unresolved identities, evidence, conflicts and publication decisions remain private; raw conversations do not become a second search index.
The memory lifecycle
Seven controlled transitions, not one opaque model call.
Bound the input, resolve identity, qualify what is known, publish safely, retrieve by scope, and keep correction possible.
- 01
Authorize
Bound the source set.
Begin with sources the participant is authorized to use; availability and retention vary by deployment. - 02
Contribute
Extract replaceable contributions.
Each supported input contributes entities, claims, relationships, evidence and temporal bounds. - 03
Resolve
Connect mentions to canonical identity.
Grounded identifiers and compatible support can resolve an entity; a matching name alone cannot. - 04
Qualify
Keep status, time and conflict explicit.
Current, historical, superseded, inferred, uncertain and conflicting memory remain distinguishable. - 05
Publish
Build inspectable canonical memory.
Supported context is projected into readable memory while raw sources and private workflow state stay outside routine retrieval. - 06
Retrieve
Return a bounded context card.
The subject, agent, permitted scope and task determine which already-published context is returned. - 07
Correct
Revise without erasing lineage.
Provenance supports correction and source replacement; the complete customer-facing correction journey remains beta work.
Source availability
Authorize real sources; do not assume a universal connector catalog.
Current evidence supports specific retained conversation, explicit capture, document, trace and wiki-source paths. Calendar, CRM, support, email and product-usage availability varies by deployment.
Reachability does not make credentials, regulated material or unusually sensitive data acceptable.
Qualification and control
State tells the agent what it may use—and how.
Time, support, conflict, correction, scope and authority travel with the context instead of hiding behind a confident answer.
| State or control | Meaning | Agent treatment |
|---|---|---|
| Current | Supported and presently applicable. | Use only within the authorized scope and task. |
| Historical | Previously true or relevant. | Use to explain change, not as the current instruction. |
| Superseded | Replaced by newer supported context. | Prefer the replacement while preserving lineage. |
| Inferred or uncertain | An evidence-bounded interpretation or unresolved state. | Qualify it and seek confirmation when consequential. |
| Conflicting | Active support disagrees. | Surface the conflict; do not silently choose a story. |
| Correction and replacement | Operator controls and source generations can revise published memory. | Current ledger mechanics; the complete self-service customer journey remains beta. |
| Scope and authority | Current retrieval is bounded inside the verified deployment. | Broader personal and organizational isolation and selected sharing remain product direction. Memory never authorizes external action. |
Closed beta
Evaluate the memory before trusting the answer.
Review identity, evidence, freshness, conflicts and retrieved scope—not only whether the final response sounds plausible.