Memory System
Thallus learns from your conversations and builds a persistent memory of your professional context. This makes future responses more relevant and personalized over time.
What Thallus remembers
After each conversation, Thallus analyzes the exchange and extracts professionally relevant information — your role, team, preferences, and recurring patterns. These memories are stored per-user and recalled in future conversations to provide better context to the planner and agents.
Thallus only stores work-relevant memories. Personal details, casual remarks, and low-importance observations are filtered out.
Memory categories
Memories are organized into four categories:
| Category | What it captures | Example |
|---|---|---|
| Memories | General professional context | "Worked on the Q4 product launch with the marketing team" |
| Preferences | Work style and communication preferences | "Prefers concise bullet-point summaries over narrative responses" |
| Facts | Role, team, projects, organizational structure | "Senior Product Manager on the Growth team, reports to VP of Product" |
| Patterns | Recurring work habits and workflows | "Reviews sales dashboards every Monday morning" |
How memories are extracted
After a conversation, an LLM analyzes the exchange and identifies potential memories. The extraction is focused on enterprise-relevant information:
- What gets extracted — Role details, team structure, recurring questions, stated preferences, project context
- What doesn't — Casual chit-chat, one-off questions with no lasting relevance, personal information unrelated to work
Memories are filtered for professional relevance before being stored.
Importance scoring
Every memory has an importance score on a 0.0–1.0 scale:
| Score Range | Meaning | Example |
|---|---|---|
| 0.9–1.0 | Core identity/role information | "VP of Engineering at Acme Corp" |
| 0.8–0.9 | Enduring preferences and key context | "Always include data sources in reports" |
| 0.7–0.8 | Useful professional context | "Currently focused on reducing customer churn" |
| Below threshold | Not stored | Filtered out during extraction |
Higher-importance memories have more influence on future responses and are retained longer.
Deduplication and boosting
Thallus automatically deduplicates memories, reinforcing important information over time. Frequently reinforced information (like your role or team) naturally rises in importance, while one-off details stay at their initial level.
How memories are used
When you send a message, Thallus retrieves relevant memories and injects them into the planning and agent prompts. Only memories relevant to your current query are retrieved — not everything Thallus knows about you, just what's pertinent.
Retrieved memories are formatted as structured context and provided to both the planner (which uses them to design better plans) and individual agents (which use them to tailor their responses). For example, if Thallus remembers you prefer bullet-point summaries, the synthesis will favor that format.
Managing your memories
You can view and manage your memories from the Settings page:
From the memory management page, you can:
- View all your stored memories with their category, importance, and creation date
- Filter by category to see only Facts, Preferences, etc.
- Edit a memory's content to correct inaccuracies
- Delete individual memories you don't want Thallus to use
- Create memories manually if you want to explicitly tell Thallus something about yourself
Privacy and deletion
Memories are scoped entirely to your user account. No one else in your organization can see your memories, and they are never shared across users.
If you delete your account, all associated memories are permanently deleted as part of the cascade deletion process. You can also selectively delete individual memories at any time from the Settings page.
For more on how memories influence the orchestration pipeline, see How Orchestration Works. For how different chat modes use memory context, all modes inject memories into their prompts, but Research and Investigate modes benefit most since they use memory during both planning and agent execution.