Frona gives agents user-scoped memory that persists across conversations. A remembered preference, project, person, or procedure can be available to every agent acting for the same user without being copied into every chat.
Memory backends
Frona has two memory backends:
| Backend | How it works | Best for |
|---|---|---|
| Basic | Stores direct user and agent memory entries, then periodically compacts them into summaries. | Existing installations that want the original lightweight memory behavior. |
| PKM | Consolidates conversations into grounded atomic memories, typed entities, relationships, playbooks, and readable Markdown pages backed by an ontology. | Inspectable long-term personal knowledge with provenance and structured retrieval. |
Fresh installations choose a backend during setup. Existing installations remain on Basic until an administrator explicitly enables PKM. See Personal Knowledge Management for the complete PKM design and workflow.
Remembering something immediately
Tell an agent naturally:
- “Remember that I prefer dark mode.”
- “Keep in mind that our team uses Slack.”
- “Remember the staging deployment steps we just used.”
The agent can call its memory tool during the conversation. With PKM, that creates a short memory immediately, so the information is available before the background consolidation pipeline has processed the chat.
What durable memory contains
Basic memory
Basic memory maintains compact user facts shared across agents and private notes scoped to an individual agent. Older entries are periodically summarized to control prompt size.
Personal knowledge management
PKM builds a user-scoped knowledge graph from several kinds of durable material:
- Entities. Identified subjects such as people, organizations, projects, services, and files.
- Atomic memories. Individual claims retained with source evidence and lifecycle history.
- Relationships. Asserted or inferred typed links between entities.
- Attributes. Structured properties defined through the effective ontology.
- Playbooks. Reusable procedures grounded in successful workflows and tool executions.
- Wiki pages. Human-readable Markdown projections generated from the structured state.
Correcting memory
Tell the agent what changed:
- “Actually, we switched from Slack to Discord.”
- “Update my timezone; I moved to Eastern time.”
- “The production database now listens on port 5433.”
Basic memory updates its compacted record. PKM preserves both the earlier and newer atomic memories, records how they relate, and projects only the current knowledge as authoritative. History and supporting evidence remain inspectable.
How agents use memory
Memory is injected selectively rather than loading every stored item into every prompt:
- Small, high-value context such as profile information and recent memory is pushed automatically.
- PKM supplies a compact index of available playbooks.
- Agents search for relevant knowledge pages when a task needs more detail.
- Full Markdown pages are read only on demand.
- Useful pages are cited, contributing a usage signal to later retrieval.
This keeps prompts focused while still letting an agent recover knowledge it did not know to ask you to repeat.
Managing memory
With PKM enabled:
- Open Memory from the user menu to browse the knowledge graph, search pages, inspect ontology structure, and trace atomic memories to their evidence.
- Open My Settings → Memory to see the active consolidation run and recent processing history.
- Use Reset memory only when you intend to delete derived PKM state and rebuild it from saved chats and retained short memories. Back up the instance first.
Administrators select the backend and background model in Administration → Memory. PKM uses the memory model group when present and otherwise falls back to primary.
Tips
- Be explicit when something should be remembered immediately.
- Correct outdated knowledge instead of asking the agent to forget the entire subject; PKM can preserve the history while changing what is current.
- Use files for large datasets and source documents. Memory should retain the durable knowledge and references, not duplicate an entire corpus into prompts.
- Check the Memory browser when you want to understand why an agent believes something.
Next steps
- Personal Knowledge Management. How grounding, ontology validation, consolidation, and retrieval work.
- Chatting with Agents. How memory participates in conversations.
- Agent Workspaces. Persistent files agents can read and create.
- Config File → Memory. Backend and consolidation settings.