Skip to content

Knowledge and memory architecture

Agents need more than a context window. They need memory, grounded knowledge, and a model of the world. These are different layers with different failure modes.

Four layers

Layer Question AxisVela examples
Context What is in front of the model right now? Prompt assembly, retrieved passages, session state
Memory What should persist across runs? MNEMOVELA: typed records, branches, retrieval, forgetting
Grounding What do terms and claims mean, and what evidence supports them? GNOSIVELA: concepts, entities, assertions, conflicts, scoped views
World model What is true about the operating environment, and when? ONTOVELA: bitemporal state, assertions, twin snapshots

A vector store is none of these by itself. It is an index. Memory, grounding, and world models require structure, lifecycle, and provenance.

Memory architecture

A useful memory design answers:

  • What is stored? Episodic events, semantic facts, procedural knowledge, user preferences, decisions.
  • How is it addressed? By identity, tenant, session, entity, time, or purpose.
  • How is it retrieved? Keyword, vector, graph, recency, importance, or a combination.
  • How is it revised? Append, branch, merge, supersede, forget.
  • How is it governed? Retention, consent, access control, deletion.

MNEMOVELA’s design emphasizes typed records, Git-like branches, hybrid retrieval, and governed forgetting. Those are architectural choices, not feature checkboxes.

Grounding architecture

Grounding is the bridge between retrieval and meaning:

  • concepts have definitions;
  • entities have stable identities;
  • claims carry evidence;
  • conflicts are explicit;
  • views are scoped to a purpose.

GNOSIVELA calls the output a grounding view. The important property is that a consumer can see not just the context, but why it should trust it.

World-model architecture

An operational world model must be temporal and evidence-bearing:

  • valid time — when something was true;
  • recorded time — when the system learned it;
  • assertions — who claimed it, with what evidence;
  • snapshots and changes — what the state was at a point in time.

ONTOVELA provides this layer for twins and operational state. PEIRAVELA uses world models as the starting point for simulation.

Design rules

  1. Separate context from memory. A long prompt is not durable memory.
  2. Separate memory from knowledge. Remembering a claim is not the same as grounding it.
  3. Separate knowledge from world state. A definition is not an observation.
  4. Carry provenance. Every retrieved item should be traceable.
  5. Make freshness explicit. Stale context is a silent failure.
  6. Govern deletion. Forgetting is a requirement, not an edge case.
  7. Scope by purpose. A planner, a support agent, and an auditor may need different views of the same knowledge.

Anti-patterns

  • “We put everything in a vector database.” That is an index, not an architecture.
  • “The model will figure out the definitions.” It will guess.
  • “The latest row is the truth.” It may be the latest belief, not the truth.
  • “We never delete anything.” That is a compliance and quality problem.

What to take away

  • Context, memory, grounding, and world state are distinct layers.
  • Provenance and lifecycle are part of the architecture.
  • Retrieval quality is necessary but not sufficient; grounding and temporality are what make knowledge usable for decisions.