Manage memory and current context
Use long-term memory for information you want available beyond a single conversation. Use the current task brief for instructions and evidence that matter right now. Neither is a substitute for checking the source of an important fact.
On this page
These controls describe the October 2026 development interface. Your installed release may differ; this is not a claim that every control is available in v0.0.1.
Know what you are saving
- Conversation history records the exchange in a session.
- Active context is the material assembled for a model request. Long conversations can be summarized; a visible history is not a promise that every message is included verbatim.
- Long-term memory holds information that can be used across sessions, with a type, scope and source evidence.
- A skill describes a repeatable method. Save a procedure as a skill rather than filling memory with a whole manual.
Memory can come from a manual entry, an explicit request to remember something, or background extraction from conversation history. Do not assume that only statements beginning “remember” can contribute to memory. Extraction is selective, and a saved entry is not guaranteed to be selected for every future turn.
Inspect what is remembered
- Open My memories.
- Search for the relevant subject. Use the type, scope and source filters to narrow the list.
- Open View details on an entry.
- Read its content and Original evidence, not just the title. Check when it was recorded, its scope and whether the source still supports it.
- If you intend to rely on it for a current decision, confirm that it is still applicable.
A confidence label describes the record, not an independent fact-check. A remembered deadline or decision may have changed since its source conversation.
Add a deliberate memory
Use Add memory to record a stable, non-sensitive fact or preference. Select a type, give it an appropriate scope, and write a short content statement. Use the title and context fields to make its meaning clear.
For example, if it reflects your actual preference: “For my weekly project briefs, put unresolved decisions before background notes.” This is a preference, not authorization to publish or send a brief.
Submit the form, then confirm the entry appears and inspect its details. If you ask the agent to remember something instead, the change can be queued for background processing. A conversational acknowledgment alone does not prove it has been stored; check My memories.
Do not save passwords, API keys, authentication tokens or other secrets. Even where the product checks for credentials, do not rely on detection to make unsafe input safe.
Correct stale information
If an entry is wrong, stop relying on it. The memory page provides add, inspect and delete controls—not a general inline editor. You can remove the incorrect entry through its delete action and add a corrected entry when appropriate. Read the confirmation carefully: deletion is not an undoable editing shortcut.
Check the list after the operation. This removes the selected memory record; it does not demonstrate that source conversations, exported copies, backups or provider-side records have been erased. Clearing a session's history is also not the same operation as deleting long-term memory.
Keep the current task explicit
For important work, state the current constraint directly: “Use the attached approved plan dated Tuesday; do not use an older launch date.” If a response appears to rely on stale information, identify the conflict and ask it to recheck the source. Memory can reduce repeated explanation, but it cannot replace fresh evidence or grant new permission.
Read sessions and context for conversation history, or skills for repeatable procedures.