> ## Documentation Index
> Fetch the complete documentation index at: https://docs.opengeni.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory and knowledge

> What agents remember, what they can look up, and what tells them how to behave.

OpenGeni keeps three kinds of durable context apart, because they carry different authority and follow different lifecycles.

## Workspace Memory

Specific facts, decisions, incidents, fixes, and confirmed outcomes. When Memory is enabled for a workspace, live agents save these autonomously during a run with `memory_save`, correct them with `memory_correct`, and find them later with `memory_search`. Saved text is stored exactly as written and is visible in the workspace's Memory view.

Memory is retrieval context, never conversation history: the model sees relevant memories, not a transcript of how they were created.

Integrations can choose a per-session `memoryScope`: `workspace`, `user`, `session`, or `off`. User and session scopes can read shared workspace facts, but their saves, corrections, and deletions stay within their own memory layer. Reading a shared fact does not allow a private chat to change it for everyone.

The [chat integration](/guides/integrate-your-product#choose-agent-access-and-memory) exposes this as `memory: "workspace" | "user" | "session" | false`. If omitted, memory follows the chat's `agentAccess` scope, which defaults to `"session"`. User memory follows the product's opaque user label within the tenant and requires that label. `false` disables Memory tools, not the conversation's durable history, and it does not change which other sessions the agent may reach. Memory retrieval and saving also require Memory to be enabled for the workspace.

## Documents and Knowledge

Upload documents or sync them from connected sources such as Google Drive. OpenGeni indexes them for hybrid vector and keyword search, and agents retrieve permission-filtered passages with citations. A separate curated Knowledge lane holds human-reviewed claims for facts that should not change without review.

## Behavior: instructions and Skills

* **Workspace instructions** are unconditional rules every agent must follow. They are composed into the prompt of every session, so keep them short.
* **Skills** are reusable, conditional how-to guidance the agent loads when relevant.
* **Organization profile** holds the organization's identity and mission and is composed ahead of workspace rules.

The same material should not be copied across these. A workspace instruction can say "search Memory for related incidents"; the incident itself stays in Memory and the investigation method stays in a Skill.

<Note>
  Skills, Packs, and the capability catalog are evolving quickly. The current contracts live in the
  repository under
  [docs/capabilities.md](https://github.com/Cloudgeni-ai/opengeni/blob/main/docs/capabilities.md)
  and [docs/packs.md](https://github.com/Cloudgeni-ai/opengeni/blob/main/docs/packs.md).
</Note>

## Learning mode

A workspace learning policy governs whether agents may derive behavior changes from what they observe:

| Mode         | Effect                                                                                 |
| ------------ | -------------------------------------------------------------------------------------- |
| Off          | No derived proposals                                                                   |
| Review first | Agents leave auditable, inactive proposals for a human to activate                     |
| Autonomous   | Eligible Skills and instruction policies activate through their own undoable lifecycle |

Learning mode does not gate Memory. Agents save and correct facts whenever Memory is enabled.