Skip to content

Memory

Memory is how Pulse remembers what matters about you and your work beyond a single exchange. It is a structured store of facts the assistant can search, update, and recall, separate from the raw timeline of messages.

Pulse memory settings showing memory metrics, search, lifecycle tabs, and stored facts

Facts as the unit of memory

Memory holds facts. Each fact is a discrete piece of durable knowledge the agent has learned — a preference, a project detail, a relationship, a decision. Storing knowledge as facts lets the agent recall precisely what is relevant rather than re-reading the entire history.

The agent retrieves memory through semantic search. Instead of exact keyword matching, it finds facts by meaning, so a question about "the launch" surfaces facts about the relevant release even if the wording differs.

Each fact carries a confidence score, signaling how strongly the agent trusts it. Confidence informs how the agent weighs a fact when reasoning and when deciding whether to act on it.

Write policies

How facts get written is governed by a write policy:

PolicyBehavior
AutoThe agent writes facts as it learns them, without prompting
ManualFacts are recorded only when you direct it

Auto keeps memory current with no effort; manual gives you tight control over what the agent retains. You choose the policy that fits the workspace from settings.

Salience-based maintenance

Memory is maintained over time using salience. As facts accumulate, the most relevant and important ones are kept prominent, while stale or low-salience facts recede. This keeps the store useful rather than letting it grow into noise.

Memory and context

Relevant facts are folded into Pulse's working context as the timeline grows. This is what lets durable knowledge survive even when the original message is no longer near the latest part of the thread.

Managing memory

The memory area in settings lets you review stored facts, adjust the write policy, and curate what the agent knows. Because facts are explicit and searchable, you can see exactly what informs the agent's behavior and correct anything that is wrong.

Studio · Pulse — Cognipeer product documentation