Conversational memory

Compass

What a system carries forward between turns, and what it deliberately forgets.

What is conversational memory?

Conversational memory is what a system carries forward between turns of a dialogue. It holds the entities already named and the constraints already stated, so a follow-up question can be interpreted without repeating them. It also determines what is deliberately discarded, since retained material can both mislead a later answer and outlive its permission to exist.

Also called: Session memory, Dialogue memory

Why what a system forgets is a design decision too

The complaint that reaches a support queue is rarely that the system forgot something. It is that the system remembered a constraint from an earlier turn and quietly applied it to a question where it no longer held. Memory also carries a retention obligation, because a dialogue in a regulated setting may contain material the organisation is required to delete on request. Deciding what persists, and for how long, belongs in the design rather than in the session store's default behaviour.

  • A follow-up question is unanswerable without the entities from the previous turn.
  • Stale constraints carried forward produce answers that were correct two turns ago.
  • Retained dialogue inherits the retention and deletion duties of any other record.
  • Memory shared across users is a disclosure route rather than a convenience.

How conversational memory works

  • Each turn records the entities and constraints it introduced, separately from the text.
  • Follow-up questions are rewritten into standalone form before retrieval occurs.
  • Constraints carry an explicit scope, so they expire rather than persisting indefinitely.
  • Memory is partitioned per identity, with no state shared between users.
  • The retained session is subject to the same retention schedule as the source records.
  • A user can inspect and clear what the session holds about them.

How conversational memory is evaluated

Conversational memory is assessed on whether a follow-up is understood and whether anything stale survives. Examine the following on real multi-turn sessions:

  • accuracy on follow-up questions that depend on an earlier turn
  • frequency of a stale constraint applied to a later question
  • correctness of the standalone rewrite against a human reference
  • retention period of session state against the applicable policy
  • leakage of state between identities under concurrent sessions

Limitations, risks and common failure modes

  • Pronoun resolution attaches a follow-up to the wrong entity from two turns earlier.
  • Permissions are evaluated on the first turn and inherited by every later one.
  • Session stores default to indefinite retention regardless of the material inside.
  • A summarised history drops the one constraint the user cares about.
  • Long sessions accumulate history that displaces the evidence for the current question.

All 315 terms