Similarity Is Not Memory
Vector retrieval has become the common way to give AI systems access to information. It is highly effective at answering one question: what stored information resembles this request?
Resemblance, however, is only one component of memory. Two passages may use similar language while referring to different people, different time periods, or contradictory claims. Two others may share no vocabulary at all while being connected through a shared entity, event, or causal relationship.
Durable machine memory must represent more than semantic proximity.
A Layered Model of Memory
We're researching memory systems that combine four complementary capabilities:
- Vector representations provide associative recall and graded semantic similarity.
- Graphs preserve identities, relationships, events, evidence, and temporal structure.
- Ontologies define stable concepts, allowable relationships, and logical constraints.
- Local AI models perform private, repeatable extraction, classification, and concept proposal close to the source data.
Each answers a different question. Vectors: what resembles this? Graphs: what is connected, and how? Ontologies: what kinds of things and relationships are valid? Provenance: why should the system believe this?
The goal is not to force every form of memory into one representation. It's to maintain a coherent logical memory while letting different physical representations do the work they're best suited to.
From Retrieval to Concept Formation
A more capable memory system should do more than return previously stored text. It should observe recurring entities, relationships, and patterns across many sources and over time — which opens the possibility of machine concept formation.
A repeated phrase should not automatically become a permanent concept. Candidates must be evaluated for recurrence, stability, usefulness, consistency, and explanatory value. They may be promoted, merged, specialized, retained provisionally, or rejected as further evidence arrives.
The distinction is fundamental: stable concepts should persist, while transient details stay tied to their original episodes and time periods.
Research Questions
- Can graph-guided retrieval outperform vector retrieval on relational and multi-step questions?
- Do ontology constraints reduce invalid entity merges and unsupported relationships?
- Can smaller local models support reproducible ontology induction?
- How should a memory system distinguish observations, inferences, disputes, and accepted beliefs?
- Does consolidation improve concept stability as new information arrives?
- How do provenance and counterevidence affect unsupported AI-generated answers?
- How should forgetting, supersession, and belief revision operate without destroying historical evidence?
Why Local AI Matters
Many of the most useful memory sources contain private documents, communications, operational records, or regulated data. Local AI allows selected extraction and consolidation to happen inside infrastructure the data owner controls.
Local operation doesn't automatically guarantee privacy or accuracy — models, embeddings, derived graphs, and logs all still need governance. But it creates real opportunities for control, reproducibility, lower latency, and the use of sensitive information that shouldn't be sent to a broadly hosted service.
Why It Matters
A useful machine memory should associate, relate, generalize, remember time, preserve evidence, and revise itself when the world changes.
By combining vectors, graphs, ontologies, provenance, and local AI, this research aims to move past retrieving relevant text toward maintaining an inspectable, evolving model of what a system has observed, what it currently believes, and why those beliefs should be trusted.