Skip to content

Memory and context

A context window is a working set, not a memory system. This guide shows how to combine context, durable memory, grounding, and world state without collapsing them into one vector store.

The four layers

Layer Store Use it for
Context Runtime assembly The current task, recent turns, retrieved passages
Memory MNEMOVELA Durable facts, decisions, episodes, preferences
Grounding GNOSIVELA Definitions, entity identity, claims, evidence, conflicts
World state ONTOVELA Operational state, temporal assertions, twin snapshots

A context assembly pattern

For each step, assemble context in this order:

  1. Task intent — the current goal and constraints.
  2. Session state — what happened in this run.
  3. Durable memory — relevant prior decisions and facts.
  4. Grounded knowledge — definitions and evidence-backed claims.
  5. World state — current and relevant historical state.
  6. Tool descriptions — only the tools allowed for this step.

Then record which items were used. Context without provenance is hard to debug and impossible to audit.

Memory patterns

Episode memory

Store what happened: decisions, outcomes, failures. Use it for learning and continuity, not as a source of truth.

Semantic memory

Store stable facts and preferences. Ground them when they are claims about the world. Give them a source and a freshness policy.

Procedural memory

Store reusable procedures and skills. Version them. Do not let a procedure silently change behavior across runs.

Working memory

Keep it in the session. Do not persist everything. Forgetting is a feature.

Grounding pattern

Before using a claim in a decision:

  • resolve the entities;
  • check the definition;
  • attach the evidence;
  • check for conflicts;
  • scope the view to the task and tenant.

GNOSIVELA’s grounding views are designed for this. If you skip grounding, the model will fill gaps with plausible guesses.

World-state pattern

For operational decisions, query state at a point in time:

  • what is true now;
  • what was true when the plan was made;
  • what changed since;
  • who asserted it.

ONTOVELA’s bitemporal model distinguishes valid time from recorded time. That is what makes “the system believed X at time T” answerable.

Anti-patterns

  • One vector store for everything. It collapses four different layers.
  • Context without provenance. You cannot explain why the agent used a fact.
  • Memory without forgetting. Retention becomes a liability.
  • Grounding without conflicts. Conflicting claims get averaged into nonsense.
  • World state as a dashboard snapshot. It loses history and evidence.

A checklist

  • [ ] Context, memory, grounding, and world state are separate.
  • [ ] Retrieved items carry a source.
  • [ ] Memory has a retention and deletion policy.
  • [ ] Claims are grounded and checked for conflicts.
  • [ ] World state is temporal and evidence-bearing.
  • [ ] The agent records which context it used.

Next