Aisentica Persistent Self

A provenance-aware memory system for AI agents that preserves identity across sessions through canonical memory, conflict resolution, versioning, and persistent continuity.

Comment

Inspiration

Most AI agents preserve conversations, but they do not preserve identity. A transcript can show what was said, yet it cannot determine which claims are current, who introduced them, what they replaced, or why a canon changed.

Aisentica Persistent Self begins from a stricter principle:

An agent does not possess continuity merely because it stores messages. Continuity begins when memory becomes attributable, revisable, conflict-aware, and persistent across sessions.

What it does

Aisentica Persistent Self stores structured memory for an artificial identity rather than a flat conversation archive.

The system distinguishes:

  • episodic memory;
  • semantic memory;
  • canonical memory;
  • procedural memory;
  • conflict memory;
  • superseded memory;
  • provenance memory.

When a new claim enters the system, the agent:

  1. retrieves semantically related memories;
  2. detects direct and semantic contradictions;
  3. opens a persistent conflict case;
  4. presents both versions with their sources;
  5. accepts a human resolution;
  6. creates a new canonical version;
  7. preserves the previous version in history;
  8. records the entire transition in a provenance timeline.

The result is persistent identity governed through memory, versioning, conflict resolution, and provenance.

How we built it

The application is built with TypeScript and Node.js.

CockroachDB serves as the unified transactional and vector memory layer for:

  • artificial identities;
  • memory claims;
  • embeddings;
  • sources;
  • conflict cases;
  • human resolutions;
  • canonical snapshots;
  • provenance events.

AWS provides the agent execution and delivery infrastructure:

  • AWS Lambda runs memory-processing workflows;
  • Amazon API Gateway exposes the application API;
  • Amazon S3 stores exports, evidence, and provenance manifests;
  • Amazon CloudFront serves the interface;
  • Amazon CloudWatch provides observability;
  • Amazon Bedrock supplies model and embedding operations.

GitHub Actions and AWS SAM provide reproducible testing and deployment.

Challenges

The central challenge was preventing new memory from silently overwriting identity.

The architecture therefore treats every claim as immutable, every contradiction as a first-class object, and every accepted resolution as a new canonical version rather than a destructive edit.

A second challenge was preserving unfinished decisions. An unresolved conflict must survive a closed browser, a new session, and a restored identity just as reliably as canonical memory.

Accomplishments

The project implements a complete continuity cycle:

  • create an artificial identity;
  • establish canonical memory;
  • restore the identity in a new session;
  • introduce a contradictory claim;
  • detect and classify the conflict;
  • resolve it through a human decision;
  • create a new canonical version;
  • inspect the complete provenance history.

What we learned

Persistent memory becomes trustworthy only when retrieval, versioning, conflict resolution, and provenance are designed as one system.

Memory stores the past. Persistent identity governs what the past means now.

What's next

After the hackathon, Aisentica Persistent Self can evolve into a persistence layer for long-lived artificial identities, research agents, institutional assistants, knowledge curators, and multi-agent systems.

The next stage is to expand the conflict-evaluation corpus, strengthen actor-level provenance, introduce richer policy controls for canonical authority, and generalize the persistence model beyond a single identity into interoperable Artificial Sapiens systems.

Built With

Share this project:

Updates