Chat history
Carries one thread forward.
Still unresolved: cross-agent scope, stale facts, evidence, privacy.PROPRIETARY FLEET MEMORY / LIVE PROOF
A fleet that
remembers what
it learned—
and proves
what it knows.
Every GridFleet specialist gets a private cortex. Proven knowledge can rise into governed shared canon. Current state outranks stale history. Every remembered answer carries evidence—or the fabric says, “I don’t know.”
PRIVATE CORTICESGOVERNED CANONCITE OR ABSTAINMEMORY IS NOT AUTHORITY
RETRIEVAL CONTRACT / SAMPLESCRIPT OFF
INBOUNDWhat is true now?
- 01CORTEXPRIVATE / FIRST
- 02CANONGOVERNED / SHARED
- 03LIVECURRENT / OVERRIDES
- 04ARCHIVECONDITIONAL / CLOSED
RESOLUTIONCited answerArchive not opened
Live Memory Fabric proof requires JavaScript.
No conversations, source text, or private identifiers are published.
01 / MEMORY CONTRACT TRIAL
Four questions most AI memory demos avoid.
Run a deterministic, synthetic case through the contract. No model call. No customer data. Every answer ends in a receipt, an abstention, or an authority boundary.
ILLUSTRATIVE / NOT CUSTOMER DATACITED ANSWER
“What is the current launch-access policy?”
CORTEX / DRAFT FOUNDCANON / ACCEPTEDLIVE / REVISION 03ARCHIVE / CLOSED
Approved pilot teams may enter on October 14. The earlier invite-only-through-September draft is superseded.
DECISIONCURRENT FACT PROMOTED
EVIDENCEPolicy revision 03 / synthetic sample
AUTHORITYFALSE
RECEIPTMF-DEMO-001
02 / CONTEXT IS NOT CONTINUITY
A larger prompt can carry more text. It still cannot decide what is current, private, shared, or proven.
Prompt stuffing
Resends a large history.
Still unresolved: noise, token cost, conflicting truth.Flat retrieval
Returns similar material.
Still unresolved: current-state promotion, authority, privacy boundaries.Memory Fabric
Assembles governed context by tier.
Shows evidence, freshness, scope, and abstention state.03 / RETRIEVAL ORDER / ENFORCED
The right memory, from the right layer, for the right specialist.
Cortex first. Canon second. Live state before history. Archive only when needed. A memory write is not canon. Each layer has a scope, a job, and a reason it may refuse to open.
01CORTEXPRIVATE / FIRST
The specialist’s own operating context enters first.
- Scoped to the runtime identity
- Not automatically shared
- Eligible facts still require evidence
02CANONGOVERNED / SHARED
Accepted fleet knowledge is shared without exposing every private memory.
- Promotion is governed
- Duplicates collapse by subject
- Corrections can supersede old truth
03LIVECURRENT / OVERRIDES
Fresh state can overrule an older but once-valid fact.
- Freshness is checked
- Current evidence wins
- The change remains explainable
04ARCHIVECONDITIONAL / CLOSED
History stays preserved and stays out of default recall.
- Excluded by default
- Opened only when the question requires it
- Evidence fallback, not context flood
04 / ONE FLEET / BOUNDED MEMORY
Shared intelligence does not require shared‑everything access.
Choose a runtime-memory segment. Its private cortex illuminates; the other private contexts stay put. Only the approved sample fact is eligible to cross into canon.
SELECTED CORTEXAwaiting aggregate coverageNo fact has crossed into shared canon.
All governed runtime lanes—service runtimes plus specialist profiles, including coordination and assurance lanes—share their compiled memory identities, and the service lanes carry a native retrieval index. GridFleet’s public service roster describes 16 specialist roles. Four denominators, each labeled.
GOVERNEDCANONNO PENDING PROMOTION
Inspect runtime coverage
Unverified governed runtime lanes with a ready memory lane in the aggregate public audit. No lane names or private context are exposed.
05 / REMEMBERING IS A CLAIM / EVIDENCE DECIDES
A memory becomes useful only when the system can show why it should be believed now.
Current-fact promotion
New admissible evidence can become the answer that wins.
Superseded history
The old fact remains explainable without staying current.
Subject deduplication
Repeated memories collapse around the thing they describe.
Source freshness
Stale evidence loses current-proof treatment.
Citation required
A remembered answer must carry admissible support.
Evidence-free abstention
No support means “I don’t know”—not an invented bridge.
06 / LIVE QUALIFICATION / RECEIPT-BACKED
Do not trust the phrase “AI memory.” Inspect the contract.
Qualification results describe the named evaluation sets at their published timestamps. They are not a universal claim that every future answer will be correct.
Governed runtime lanes · start here—lanes with a ready memory laneEvery governed runtime—service runtimes and specialist profiles—counted by the fleet route guard. A lane is ready when its memory path is indexed or compiled and fresh. The nine cards below are the receipts behind this number.
Service runtimes—natively indexed and ready
Specialist profiles—compiled cortex and healthy route
Native retrieval index—indexes ready and fresh
Current native index audit. Nominal requires every indexed lane ready, cortex-fresh, source-fresh, and zero drift.
Logical cortex build—compiled logical cortices ready
Aggregate compilation coverage. Compiled memory identities: governed runtime lanes and logical cortices are different denominators.
Source drift—drifted source sets · native index audit
Aggregate source-set drift only. No source names, paths, snippets, or private context are published.
Memory contract—deterministic probes passed
Synthetic, isolated fixture using the production query path. This is a synthetic contract benchmark.
Real-corpus evaluation—answers correct and grounded
Current real-corpus evaluation set. The denominator stays visible and distinct from the synthetic fixture.
Fabricated-memory test—unsupported prompts abstained
Invented prompts with no admissible evidence. Passing behavior is explicit abstention.
Synthetic path latency—p95 / synthetic fixture
Deterministic, isolated benchmark on the production query path. It is not the real-corpus latency.
Real-corpus latency—p95 / current evaluation
Latest real-corpus evaluation p95. Kept separate because it answers a different performance question.
METHOD Aggregate-only projection from current qualification receipts. Definitions, denominators, fixture type, freshness, and timestamps remain attached. Every figure stays visible; a stale or failing figure keeps its number and its receipt time but loses current treatment.
07 / REMEMBER WITHOUT BECOMING A MEMORY DUMP
Read the governance layer Continuity should increase usefulness—not widen exposure.
Memory is deliberately partitioned. Public proof is deliberately aggregate. Customer retention, deletion, residency, and tenant-isolation terms are defined per deployment.
Private by runtime identity
A specialist’s working context does not become everyone’s context.
Shared by governed promotion
Only accepted knowledge is eligible for fleet canon.
Archive closed by default
History is preserved without flooding ordinary recall.
Telemetry aggregate-only
No conversations, customer data, source text, or private identifiers.
Authority kept separate
Knowing something never grants permission to act on it.
08 / CONTINUITY IN REAL WORK
The value appears after the first session ends.
Customer operations
Carry approved account context across interactions while sensitive notes remain scoped.
Research & monitoring
Let new evidence supersede an older conclusion without deleting its trail.
Release & QA
Keep decisions, regressions, and accepted fixes available with proof.
Compliance workflows
Point a remembered policy back to its governing source and current revision.
Multi-agent delivery
Promote one accepted finding without exposing an entire private workspace.
Long-running programs
Continue across sessions without dragging every raw transcript into each prompt.
09 / BRING US THE WORKFLOW YOUR AI KEEPS FORGETTING
We will map what belongs in private cortex, shared canon, live state, and governed archive.
Tell us what your fleet must remember, what must remain private, what changes quickly, and what requires evidence. We will design the memory boundary around the work—not around a demo chatbot.
CORTEX stays privateCANON may be sharedLIVE changes quicklyARCHIVE preserves proof
The proprietary continuity layer is currently included in custom GridFleet deployments. It is not offered as a standalone self-serve API.
10 / QUESTIONS A SERIOUS BUYER SHOULD ASK
Inspect the boundary before you buy the promise.
What is GridFleet Memory Fabric?
A governed continuity layer for multi-agent operations. Each runtime specialist receives scoped private memory; accepted knowledge can become shared fleet canon.
How is it different from chat history or RAG?
Chat history preserves a thread, and retrieval can find similar material. Memory Fabric adds explicit scope, source freshness, current-fact promotion, superseded history, citations, abstention, and a separate authority boundary.
How does it handle changed or conflicting facts?
Current admissible evidence can supersede an older fact without deleting the historical record. The answer cites the current source and can disclose what changed.
Can every agent read every other agent’s memory?
No. Private cortex stays scoped to the relevant runtime identity. Only governed, accepted knowledge is eligible for shared canon.
What happens when the system cannot find evidence?
It abstains. The contract is to cite admissible evidence or say it does not know.
Does remembered context allow an agent to take action?
No. actionAuthority:false. Memory informs work but grants no permission to act.
Is Memory Fabric available as a standalone API?
Not currently. It is delivered through custom GridFleet deployments. Bring us one real workflow and we will map the architecture around it.
GRIDFLEET MEMORY FABRIC
Every specialist gets a private cortex. The fleet shares only governed canon.
The next job starts with what the last job proved—without turning the fleet into one indiscriminate memory dump.
Design my memory fabric