Skip to content

Engineering note

8DB | Your AI Remembers. Who Controls What It Keeps?

Persistent memory becomes more useful when its origin, access rules and dependencies survive with it. 8DB combines those building blocks with a checkable structural-erasure path for isolated contributions.

Published
Updated
Reading time4 minutes
AI agent memoryLong-term memory for AI agents securityAgent memory governancePersistent agent stateMemory erasure AIAgentic AI data layerPost-quantum databaseAI provenance

A support agent remembers a customer's temporary delivery address. That memory saves time tomorrow. A month later, the address has expired, a summary still contains it and a different team is using the same assistant.

Remembering was the easy part. The system now needs to decide whether the memory is current, who may recall it and which later conclusions depend on it.

8DB brings several relevant capabilities together in an omnimodal database: provenance and confidence structures, cryptographically bounded protected reads, and a structural-erasure mechanism for causally isolated contributions. Together, they give builders a foundation for memory that can be reviewed throughout its lifecycle.

The benefit is continuity with control. An agent can retain useful information while the application retains the means to inspect, restrict and retire it.

Preserve the memory and its relationships

The delivery address should remain connected to the message that supplied it, the person it concerns and any summary derived from it. An inferred address should remain distinguishable from one the customer confirmed.

8DB's provenance and typed confidence structures supply building blocks for those distinctions. Its graph and other native representations give applications ways to relate a memory to its sources and uses. The integration must record those relationships when it creates or transforms information; a database cannot reconstruct dependencies that were never supplied.

This is where a connected store can help. Retrieval, source review and change-impact analysis can work from linked information instead of relying on a loose collection of summaries. The diagram shows the lifecycle a memory application should implement using those capabilities.

8DB · Architecture in practice

Keep control through the memory lifecycle

Illustrative temporary delivery address and a summary derived from it.

  1. 01 · CaptureRetain the origin

    Link the memory to its source and record the relationships the application creates.

  2. 02 · RecallCheck current access

    Use the protected retrieval path and the application's current authorization controls.

  3. 03 · RetireInspect dependencies

    Distinguish an isolated represented contribution from one entangled with later work.

Isolated under braid rulesRemove and verify

The scoped structural-erasure path returns a certificate.

Entangled with later contributionsResolve the dependencies

The fast path refuses and returns information for further handling.

Conceptual lifecycle, not an automatic end-to-end deletion service. The erasure certificate covers the represented contribution; external copies and media sanitization need separate controls.

Check access when memory is recalled

A remembered fact may outlive the permission that originally made it available.

8DB's tested protected-read path binds decryption to a governance context. A different context can fail to open the same protected bytes. For an agent deployment, the authority that issues access contexts and controls root secrets must remain outside the agent's reach.

That boundary complements permissions on the agent's tools and outputs. OWASP recommends least-privilege access for agent systems; memory should be included in that review, alongside live documents and APIs. Our article on agent data boundaries describes an evaluation through the actual retrieval path.

Retire a memory with its dependencies in view

Removing a source record and removing its influence are different operations. The temporary address might also appear in a summary, an export or a downstream decision. A useful retirement process identifies those dependencies before declaring completion.

8DB's braid mechanism has a specific structural-erasure path. When the selected contribution is causally isolated under that representation's rules, it can be removed from the represented state and an erasure certificate produced. When it is entangled with later contributions, the fast path refuses and returns dependency information for further handling.

That refusal is useful control: it identifies when retirement requires a broader repair instead of presenting a local removal as a complete result. Tests exercise isolated erasure and certificate verification; they establish the scoped mechanism, not deletion from every possible downstream system.

The certificate concerns the represented contribution. Copies in logs, backups, exports, model training or already released plaintext require their own controls. NIST's media-sanitization guidance addresses a separate layer of this problem. Structural erasure, access revocation and storage sanitization belong in one lifecycle plan, with evidence appropriate to each.

Turn a deletion request into a verifiable workflow

A practical implementation can record four things: the memory being retired, the known dependent artifacts, the operations completed and any remaining obligations. A later worker can then continue the task from explicit state.

For the address example, that may mean removing the active memory, refreshing the derived summary and recording how an archive is handled under the applicable retention policy. Those steps are an application workflow, not a claim that 8DB automatically discovers every external copy.

The database mechanisms are valuable because they give that workflow inspectable inputs and outcomes. An operator can distinguish a completed structural removal from an access change or an unresolved dependency.

Evaluate one complete memory lifecycle

Start with a memory derived from a permitted source. Retrieve it, inspect its origin and create a dependent summary. Change the user's authority and test recall again. Then retire the source memory and inspect both the isolated and dependent cases.

This exercise tells a team more than a recall demo alone. It tests whether useful memory can survive across sessions without losing the information needed to govern it.

Tell us which memory your agent needs to keep, and when it must stop using it. One real lifecycle is enough to identify the boundaries an initial integration should prove.

Evidence scope Implemented mechanisms · Scoped erasure tests · Deployment guidance. Explore the evidence.

Continue reading

The next layer of the argument.

Engineering note

Trustworthy AI Data

8DB | A Prompt Can Ask for Your Secrets. Can the Agent Open Them?

Put your AI agent’s protected data access behind encryption. With 8DB, test which records open, which stay sealed and what happens when authority changes, independently of the model’s response to a prompt.

4 min readImplementation-reviewed · Scoped evaluation paths · Conceptual visualRead article ↗
Research

Trustworthy AI Data

Give Your AI the Evidence Behind the Answer

A useful retrieval result carries its source, uncertainty and coverage as well as its content. 8DB gives AI builders those distinctions in the data layer, so applications can decide when to answer, qualify or investigate.

4 min readImplemented data types · Source review · Illustrative retrieval exampleRead article ↗
Idea

Trustworthy AI Data

Your AI Found the Right Passage. Can It Tell Whether to Trust It?

A relevant answer can come from an expired policy or hide a conflicting source. We propose testing the evidence a retrieval system hands to an agent, so teams can see what survives the journey from stored record to answer.

6 min readEvaluation proposal · Illustrative fixture · No system scores reportedRead article ↗