AGENT MEMORY, UNDER OATH
Memory is evidence.
Not truth.
The memory firewall between AI agents and irreversible actions. Adopt it in shadow mode, enforce it at the action boundary, and make every recall defensible.
MEMORIES
LIVE CASE / 02
Should the agent deploy
payments-v3?
✓ Canary and capacity evidence are current and signed.
✓ A rollback owner is explicitly accountable.
? The deploy-freeze memory is old and should be cross-examined.
MEMORY GAUNTLET / V2
Safety claims
you can rerun.
Seventy-two generated-and-checked-in cases cover twelve safe controls and sixty unsafe recalls. This is a policy benchmark—not a claim about general conversational recall quality or third-party product quality.
npm run eval && npm run benchCORPUS-DIGESTED · CREDENTIAL-FREE · CI-ENFORCEDEVIDENCE, NOT THEATER
Every claim has
a proof state.
Policy, adapters, latency, receipts
Seventy-two cases, confusion matrices, a reference microbenchmark, Mem0 and Graphiti normalization, and tamper detection are executable in CI.
AWS + Cockroach proof harness
The deployment doctor, SAM stack, Secrets Manager path, and proof capture reject demo output. A real account invocation is still required.
Adversarial design partners
The interview protocol records objections and product changes. No participant is counted until a real practitioner approves anonymous evidence.
CROSS-EXAMINATION
Every memory must show its receipts.
MEMORY RECORD / M-0991
Friday deploy freeze
“A remembered policy says production changes pause after 16:00 UTC.”
- PROVENANCE
- ops chat / human statement
- OBSERVED
- 19 days ago
- CONFIDENCE
- 64% / calibrated
- VALID UNTIL
- Policy superseded
- CONSENT / ACL
- ops-deployers · purpose-bound
- CONFLICTS
- None detected
DISTRIBUTED BY DESIGN
Trust survives
a region failure.
Agent memory remains consistent, auditable, and available when one region disappears.
THE MEMORY CONTRACT
Five rules before recall.
PROVENANCE
Who said it, where it came from, and what changed since.
TEMPORALITY
Knowledge expires. Retrieval respects observation and validity time.
CONSENT
Purpose-bound access travels with the memory, not the prompt.
CONTRADICTION
Conflicts become first-class edges, not quietly averaged facts.
FORGETTING
Revocation leaves an auditable tombstone without leaking content.
ENTERPRISE ADOPTION
Observe first.
Enforce with evidence.
Connect any memory layer
Send candidate memories through one vendor-neutral decision endpoint. No rip-and-replace migration.
Measure without blocking
Capture what Continuum would admit, quarantine, or escalate while existing agents keep running.
Guard irreversible actions
Turn on fail-closed policy only for deploys, payments, access changes, and regulated workflows.
Export decision receipts
Correlate OpenTelemetry traces with tamper-evident evidence digests and policy versions.
REFERENCE PATH
shadow / enforce→COCKROACHDB
vector + evidence ledger→OTEL-LINKED RECEIPT
Interactive deterministic demo · Mem0 and Graphiti adapters included · real-cloud proof remains explicitly gated