Manage workspace memory

This guide shows you how to manage your workspace’s knowledge store.

Memory domains

Domains group related knowledge. Create domains for different aspects of your project:

1. Navigate to Memory in the sidebar
2. Click Create domain
3. Enter a name (e.g. “conventions”, “api-design”, “security”)
4. Add a description
5. Click Create

Domains are workspace-scoped and control access boundaries.

Add facts

Facts are knowledge units that agents can retrieve during dispatch.

1. Navigate to a memory domain
2. Click Add fact
3. Enter:
   • Topic: a short label (e.g. “error handling pattern”)
   • Content: the knowledge text (e.g. “All API errors use the Problem Details RFC 7807 format with a type field.”)
   • Confidence: how confident you are in this fact (0.0–1.0)
4. Click Save

Facts are embedded automatically for vector search. Agents retrieve relevant facts based on the conversation context — you don’t need to explicitly link facts to agents.

Supersede facts

When knowledge changes, supersede the old fact:

1. Find the outdated fact
2. Click Supersede
3. Enter the new content
4. Save

The old fact is marked as superseded and won’t appear in search results, but it’s preserved for audit trail.

Add policies

Policies are normative rules derived from facts. They constrain agent behavior.

1. Navigate to a domain
2. Click Add policy
3. Enter:
   • Rule: the normative statement (e.g. “Never use print() for logging — always use the AppLog utility.”)
   • Required role: optionally gate to a specific role (e.g. only “coder” agents see this policy)
   • Source facts: link to the facts that justify this policy
4. Click Save

Active policies are injected into agent prompts when relevant.

Control access

Memory access grants determine which agent roles can read or write to each domain.

1. Navigate to Memory → Access grants
2. For each domain, set permissions per role:

Permission	Effect
none	The role cannot see facts in this domain
read	The role can search and retrieve facts
write	The role can create, update, and supersede facts and policies

For example, you might grant:

• Coder: read access to “conventions”, write access to “api-design”
• Reviewer: read access to “conventions” and “security”
• Security: read/write access to “security”

Working memory

Each agent has a private working memory scratchpad. Agents use it to persist transient state across runs (e.g. “I’m in the middle of refactoring the auth module”).

Working memory is automatically injected into the agent’s prompt at dispatch time. You don’t manage it directly — it’s agent-owned.

Search

Control Center searches memory using a hybrid approach:

• Vector search: semantic similarity between the query and fact embeddings
• FTS search: full-text keyword matching
• Hybrid (RRF): combines both for the best results

Agents automatically search memory during dispatch via BuildMemoryContextUseCase. You can also search manually from the Memory view.

Related guides

• Create and configure an agent
• Memory and knowledge