Einstein memory gives agents long-term memory: they remember facts about
users and cases and use them in future conversations. In the Web UI memory is
represented by two screens — Memory (the review queue) and Facts — while
Einstein itself is a built-in module.
Memory: the review queue
The “Memory” screen is a queue of candidate facts. Each candidate is shown as
a card: the claim text, a “subject — predicate — object” triple and an
evidence-source chip. You can Promote a claim — it becomes a fact — or
Reject it. There is a manual refresh button.
Note that mandatory human review applies only in review mode. In auto mode
claims are promoted by policy without that step, and in off mode memory is not
collected at all. The mode is set in the “Einstein (memory)” settings category.
Facts
The “Facts” screen lets you browse stored facts (filter by user or “my facts”)
and forget an individual fact. It also has:
Dream — a memory-consolidation sweep: the agent walks the facts and
reconciles them, reporting how many scopes were swept, groups scanned and
contradictions found.
Contradictions — a list of conflicting claims to resolve.
How to use it
Work the review queue regularly: promote what’s true, reject what’s wrong — that
keeps memory from getting noisy.
Run “Dream” periodically to keep memory coherent and contradiction-free.
Use “Forget” when a fact is stale or landed by mistake.
Configuration
An agent’s memory policy is set by the memory_policy section in the
agent profile. Einstein appears on the
Plugins screen as a built-in module (no uninstall);
memory collection can still be turned off via the off mode — the module
stays in place.
**Einstein memory** gives agents long-term memory: they remember facts about
users and cases and use them in future conversations. In the Web UI memory is
represented by two screens — **Memory** (the review queue) and **Facts** — while
Einstein itself is a built-in module.
## Memory: the review queue
The **"Memory"** screen is a queue of candidate facts. Each candidate is shown as
a card: the claim text, a "subject — predicate — object" triple and an
evidence-source chip. You can **Promote** a claim — it becomes a fact — or
**Reject** it. There is a manual refresh button.
Note that mandatory human review applies only in `review` mode. In `auto` mode
claims are promoted by policy without that step, and in `off` mode memory is not
collected at all. The mode is set in the "Einstein (memory)" settings category.
## Facts
The **"Facts"** screen lets you browse stored facts (filter by user or "my facts")
and **forget** an individual fact. It also has:
- **Dream** — a memory-consolidation sweep: the agent walks the facts and
reconciles them, reporting how many scopes were swept, groups scanned and
contradictions found.
- **Contradictions** — a list of conflicting claims to resolve.
## How to use it
- Work the review queue regularly: promote what's true, reject what's wrong — that
keeps memory from getting noisy.
- Run "Dream" periodically to keep memory coherent and contradiction-free.
- Use "Forget" when a fact is stale or landed by mistake.
## Configuration
An agent's memory policy is set by the `memory_policy` section in the
[agent profile](/en/v1.0/webui/agents#behavior-sections). Einstein appears on the
[Plugins](/en/v1.0/webui/marketplace) screen as a built-in module (no uninstall);
memory **collection** can still be turned off via the `off` mode — the module
stays in place.
## Next
- [Einstein memory (concept)](/en/v1.0/concepts/memory-einstein) — how it works.
- [Knowledge base](/en/v1.0/webui/knowledge) — documents vs facts.
- [Einstein plugin](/en/v1.0/plugins/einstein) — the module's capabilities.
