Skip to content

How Memory Works

Memory lets agents retain information across conversations. When an agent learns something about you or its own context, it can store that as a fact. These facts persist and are available in future conversations.

There are several types of memory, each serving a different purpose:

Facts about you that are shared across all agents. If you tell one agent your name or preferences, any agent can access that information.

Stored with the remember_user_fact tool.

Facts specific to a single agent. These are private to the agent that created them and not visible to other agents.

Stored with the remember_agent_fact tool.

Persistent personality information for an agent. Its name, traits, communication style. Updated through the update_identity tool.

Auto-generated summaries of conversations within a space. The platform creates these automatically through compaction. Agents don’t manage them directly.

When an agent starts a conversation, the platform loads relevant memory into the context:

  1. User facts (shared across agents)
  2. Agent-specific facts
  3. Agent identity
  4. Space context (if the chat is linked to a space)

This happens automatically. You don’t need to tell the agent to “remember” to check its memory.

Agents store facts using the remember_user_fact and remember_agent_fact tools. They can also update existing facts when information changes by setting overrides: true.

The platform gives agents guidelines for memory management:

  • Read existing memory before storing new facts to avoid duplicates
  • Use overrides when correcting or updating existing information
  • Keep facts concise and specific
  • For large data, save to a workspace file and reference it in memory

Over time, memory accumulates. The platform runs periodic compaction to keep it manageable:

  • Insight compaction (default: every 2 hours). Merges, deduplicates, and cleans up memory facts
  • Space compaction (default: every hour). Summarizes conversation history within spaces

Compaction runs automatically in the background. You can tune the intervals with FRONA_SCHEDULER_INSIGHT_COMPACTION_SECS and FRONA_SCHEDULER_SPACE_COMPACTION_SECS.

See Insights for more on how compaction works.