VerifiedState: Trustworthy Memory for AI Agents

The Problem With AI Agent Memory

AI agents are being deployed into consequential workflows — customer support, legal research, medical triage, financial operations. These agents need memory. They need to remember what users told them, what decisions were made, what commitments were given.

The current state of agent memory is inadequate for this responsibility.

Most memory systems store text and retrieve it by similarity. They answer the question "what did we store that looks like this query?" They cannot answer:

• Is this fact still true?
• Where did it come from?
• Who verified it?
• Does it conflict with something else we believe?
• Can I prove to an auditor what the agent knew when it made this decision?

These are not edge cases. They are the core requirements of any memory system trusted with real decisions.

What VerifiedState Adds

VerifiedState is a memory infrastructure layer built around a single principle: every stored fact should come with proof.

Not proof in the philosophical sense — proof in the practical sense. A signed, queryable record that shows:

• What the claim is
• What evidence supports it
• How confident the system is
• Who wrote it and when
• Whether it has been challenged or superseded
• What happened when it was verified

We call this record a verification receipt. It is the core primitive of VerifiedState.

The Assertion Model

VerifiedState stores memory as assertions — structured subject-predicate-object claims rather than raw text.

Instead of storing: "The user mentioned they prefer dark mode and use Postgres"

VerifiedState stores:

• user → prefers → dark mode (confidence: 0.94, status: verified)
• user → uses → Postgres (confidence: 0.87, status: verified)

Each assertion has a full lifecycle: proposed → verified → active → superseded → retracted.

Each assertion carries temporal validity — it knows when it became true and when it stopped being true.

Each assertion is aware of conflicts — if two assertions contest the same fact, VerifiedState creates a conflict set and tracks which one has stronger evidence.

The Seven-Tier Verification Ladder

Not all facts deserve the same level of trust. VerifiedState runs every assertion through a verification ladder:

1. Structural — is the claim well-formed? Does it resolve to a valid slot?
2. Span grounding — does the cited evidence actually support the claim?
3. State consistency — does this conflict with existing verified assertions?
4. Cross-source — is it corroborated by other trusted sources?
5. Authority — is the writer authorized to make this type of claim?
6. Execution — for action claims, is there system-level proof it happened?
7. Human — escalate to a human reviewer for high-stakes or unresolvable disputes

Tiers 1-3 run on every write. Higher tiers activate based on claim class and confidence threshold.

Retrieval That Understands Intent

VerifiedState does not use a single retrieval method. It routes queries to the right channel based on what the query is actually asking:

• Exact lookup for structured queries with known subject and predicate
• Lexical search for keyword-specific queries
• Semantic search for natural language queries
• Temporal queries for time-anchored questions ("what was true last month")
• Graph traversal for multi-hop questions ("who approved the decision that caused X")
• Conflict queries for disputed facts

Results from all active channels merge and rank by relevance, verification status, recency, and confidence. The system knows when to say "I don't know" — returning a structured abstention rather than a low-confidence guess.

Proactive Memory

VerifiedState does not wait to be asked. When a fact changes — when an assertion is superseded or disputed — it notifies any agent that recently relied on that fact via webhook.

An agent acting on outdated information is not just inefficient. In consequential workflows it is dangerous. Proactive notification closes this gap.

Governance Built In

Every VerifiedState deployment includes a policy engine. Organizations define rules that govern memory behavior:

• Which agents can write which types of claims
• Which facts require human approval before an agent can act on them
• How long facts remain valid before requiring re-verification
• What gets deleted, when, and under what retention policy

Policies compose and inherit. An organization-level policy can be overridden at the namespace level, which can be overridden at the user level. Deny always beats allow.

Audit by Design

VerifiedState maintains an immutable audit log of every read, write, verification, challenge, and policy evaluation. Every receipt is signed and chained — a tamper-evident record that cannot be altered retroactively.

When an enterprise customer needs to answer "what did the agent know when it made this decision, and how did it know that?" — VerifiedState can answer completely and cryptographically.

This is not a feature added for compliance. It is the architecture.

Why Not Blockchain?

The properties VerifiedState provides — tamper evidence, signed records, audit trails — might suggest blockchain. It is the wrong tool for this problem.

Blockchain solves trust between parties who do not know each other. Enterprise AI deployments operate within a single organization's trust boundary. What they need is fast, private, cryptographically sound record-keeping — not decentralized consensus.

VerifiedState delivers cryptographic guarantees at millisecond latency, at a fraction of the cost, with complete data privacy. No wallets. No gas fees. No public ledger.

Who This Is For

VerifiedState is infrastructure. It does not have a user interface. It has an API.

The right customer is a team building an AI agent that:

• Operates in a consequential domain (support, legal, medical, financial, operations)
• Needs memory that persists across sessions
• Needs to be able to explain or audit the agent's beliefs
• Cannot afford to have agents act on stale or unverified information

If your agent just needs to remember conversation history, a simpler tool will serve you. If your agent needs to be trusted — by users, by regulators, by auditors — VerifiedState is the layer it runs on.

Current Status

VerifiedState is live. The API is available at api.verifiedstate.ai. Free tier includes 25,000 assertions, 2,000 verifications, and 10,000 queries per month with no credit card required.

Documentation: verifiedstate.ai/docs

API access: verifiedstate.ai/keys

Support: support@verifiedstate.ai