---
title: "What is cross-session memory?"
description: "Cross-session memory defined: how AI agents keep information between conversations so a returning user is recognized with full history."
canonical: https://past.dev/glossary/cross-session-memory
last-updated: 2026-09-02
---
# What is cross-session memory?

Source: https://past.dev/glossary/cross-session-memory

Cross-session memory is information that persists between conversations, so a returning user is recognized rather than treated as a stranger. It requires storage outside the model and the session: facts written during one conversation, kept with timestamps and sources, and recalled in later ones. With it, an agent asked about a customer in October can use what it learned in March, in a different session and channel.

## How cross-session memory works

1. **Capture.** During or after each session, the statements worth keeping are extracted from [session memory](/glossary/session-memory): facts, preferences, decisions, commitments. Each is stored with its event time and its source.
2. **Storage.** A [long-term memory](/glossary/long-term-memory) layer outside the model holds them: a database, a knowledge graph, or a managed memory service.
3. **Recall.** A later session queries that layer and places the results into the context window, so the model can use them as if it had been present the first time.

Recognizing that the returning user is the same person is its own problem: sessions arrive under different identifiers, and [identity resolution](/glossary/identity-resolution) links them to one profile.

## Why it matters for AI agents

Without cross-session memory, every conversation restarts. Users repeat context they already gave, agents redo research they already did, and a decision's history disappears with the session that produced it. The effect is largest in support and account work, where the relevant history is months old and spread across channels. The overview at [what is agent memory](/what-is-agent-memory) covers the layers involved.

Persistence raises requirements: a wrong fact written once can be recalled for months, so stored memory needs timestamps, sources, and a way to correct or delete entries. past.dev provides this layer as an API: text from each session goes in with its original timestamp, and recall returns current and previous facts with dated, source-attributed evidence; see the [quickstart](/docs/memory-api/quickstart).

## Cross-session memory in practice

On March 12 a customer tells a support agent that invoices must go to billing@harborlane.com from April onward. The session ends. On October 7 the same customer opens a new chat about a failed payment.

With cross-session memory, the agent recalls the March 12 instruction, checks the invoice address first, and can cite the March conversation as its source. Without it, the agent asks the customer to repeat information given seven months earlier. The value compounds across channels: the March instruction arrived in a support chat, while the October question can arrive by email and still reach the same memory.

## Related concepts

- **[Session memory](/glossary/session-memory)**: the mirror term, state that ends with the conversation.
- **[Long-term memory](/glossary/long-term-memory)**: the storage layer that persistence writes to.
- **[Identity resolution](/glossary/identity-resolution)**: linking a returning user's identifiers to one profile.
- **[Memory poisoning](/glossary/memory-poisoning)**: the security risk persistence introduces.

## Frequently asked questions

### How do AI agents remember users across sessions?

They use an external memory layer. During each session, facts worth keeping are written to it with timestamps; in later sessions the agent queries that store and loads relevant results into context.

### What is the difference between session memory and cross-session memory?

Session memory exists only inside one conversation and is discarded when it ends. Cross-session memory is deliberately persisted outside the session so later conversations can use it.

### Do AI agents have memory between sessions by default?

No. Models are stateless between calls, and most agent frameworks discard conversation state at session end. Persistence has to be added through a memory store or memory API.

## Related

- [What is session memory?](https://past.dev/glossary/session-memory)
- [What is long-term memory?](https://past.dev/glossary/long-term-memory)
- [What is identity resolution?](https://past.dev/glossary/identity-resolution)
- [Identity resolution guide](https://past.dev/guides/identity-resolution)
- [Benchmark methodology](https://past.dev/benchmarks/methodology)