# What is Short Term Memory?

Also called Conversation Memory, Session Memory.

Short term memory is the recent context an AI system keeps available within a conversation or session, typically the last several exchanges plus any state built during the current task. It lives in the prompt or in a session store, and it is discarded or summarized when the session ends.

In practice short term memory is the message history replayed into each model call. Because the context window is finite and every token costs time and money, applications trim it. Common strategies keep a fixed number of recent turns, keep a token budget worth of history, or summarize older turns into a compact recap.

The trimming boundary is where errors appear. A detail stated early, such as a chosen date or a constraint on tone, can silently fall out of the window, and the system then contradicts itself. Systems avoid this by promoting durable details to long term storage rather than relying on the buffer to hold them.

Summarization is the usual compromise. Older turns are condensed into a paragraph that preserves decisions and open questions while dropping wording. This holds continuity in bounded space but is lossy and can introduce distortions, so critical values such as identifiers, quantities, and dates are better stored verbatim in structured fields.

Terminology here is borrowed loosely from cognitive psychology and does not map cleanly onto it. In engineering usage, short term memory means session scoped context, and its boundary with working memory is drawn differently by different authors, which is a frequent source of confusion when comparing frameworks and papers.

## Key points

- Recent conversation context replayed into each model call.
- Bounded by the context window and by cost.
- Trimming can silently drop constraints stated earlier.
- Summaries preserve continuity but lose exact values.
- Session scoped, cleared or condensed when the session ends.

## In practice

A user spends ten turns planning a trip, stating in turn two that the budget is fixed and that flights must be direct. By turn thirty the earliest turns have been trimmed, and the assistant proposes a connecting flight. Promoting those two constraints into a persistent task record, and injecting them every turn, prevents the contradiction.

## Related terms

- [Working Memory](/en/glossary/working-memory)
- [Long Term Memory](/en/glossary/long-term-memory)
- [Agent Memory](/en/glossary/agent-memory)
- [Context Injection](/en/glossary/context-injection)
- [Episodic Memory](/en/glossary/episodic-memory)

[Back to the AI Glossary](/en/glossary)