# Memory

Retained state an agent can use in a later session or job — a stored preference, a prior job's outcome, or a summary extracted from an earlier interaction.

Source: https://aecon.ai/glossary/memory
Updated: 2026-09-15

## In practice

A support agent recalls that this customer prefers PDF delivery, because the customer said so and the preference was stored with a source, an owner and an expiry.

## The distinction

Memory provides continuity; it does not provide authority or truth. It is a derived record that can preserve a misunderstanding, and current source-system data always outranks a conflicting memory.

## Four kinds, four safe sources

| Type | Example | Safe source |
| --- | --- | --- |
| Session or working | The current case and steps already attempted | This session and its tool events |
| Episodic | A prior job occurred and was accepted on a date | The job and acceptance record |
| Profile or semantic | The customer prefers PDF delivery | An explicit statement or authorised profile |
| Procedural | The approved refund escalation process | A versioned operating policy |

The labels help design the system. They do not make a memory true. A generated summary is a derived record and should stay linked to the interaction it came from.

## Memory needs a correction path

People change roles, addresses and preferences. Businesses change policies and prices. An agent can also extract the wrong fact in the first place.

A system that only writes and recalls is incomplete. Consequential memories need to be visible, correctable and deletable, with the deletion following through to derived indexes. A high-impact decision should never rest on an opaque remembered inference.

## Eight questions before enabling it

What event creates a memory? Is it quoted, extracted, inferred or summarised? Who owns it? Who can retrieve, correct, export and delete it? When does it expire? Can the source and prior versions be reconstructed? What stops one user's memory reaching another user's job? And which memories are prohibited outright?

## Sources & context

- [Set up Memory Bank](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/memory-bank/set-up) — Google Cloud, 2026. Documents memory as a service attached to an agent engine instance, with configurable generation and access controlled through cloud identity. Checked 2026-09-13.
- [Managed agent memory](https://platform.claude.com/docs/en/managed-agents/memory) — Anthropic, 2026. Documents workspace-scoped memory documents with immutable versions for changes. Checked 2026-09-13.

## Related reading

- [Context](https://aecon.ai/glossary/context)
- [Retrieval](https://aecon.ai/glossary/retrieval)
- [Provenance](https://aecon.ai/glossary/provenance)
- [Agent harness](https://aecon.ai/glossary/agent-harness)