Memory management

Review what EvoX remembers, correct lasting background, and check how it contributes to later work.

Memory reduces repeated introductions and explanations. It does not mean every conversation is stored in full or every memory appears in every answer. Check three things separately: whether information was saved, whether its scope fits, and whether the current task used it.

Distinguish profile values from memory records

Open Memory and start with Memory list. This includes profile values applied to conversations and highlights recalled by relevance.

The list is a curated highlights view, not a complete export of raw records. If an item is absent, inspect recall and curation before concluding it was lost.

Put ongoing project requirements in Project working rules, and repeatable procedures in Skills. Memory supplies background and does not authorize actions.

Save information worth keeping

Keep stable, accurate facts that will remain useful. For example:

I maintain product documentation primarily for readers without a technical background.

One-time deadlines, temporary plans, and unverified judgments rarely belong in lasting memory. Do not add passwords, keys, or verification codes.

Memories accumulate through conversations, but mentioning something does not prove it became usable memory. Add important facts manually when you want to confirm them. Check facts before preserving material extracted from AI replies.

Add a memory manually

1. Find the manual-entry section in Memory list and write one concise, checkable fact.
2. Check its scope. With a current project, choose that project or all projects. Keep project-specific background out of the all-projects scope.
3. Save and wait for confirmation. Resolve any reported error before retrying.
4. Refresh the list, check content and scope, then verify retrieval in Recall test.

Without a recognized current project, a write does not automatically receive project scope. Read the scope explanation before saving.

Older entries may be unclassified. When the list offers Classify as this project or Classify as all projects, inspect the affected entries first. Project classification separates background; it is not an account permission or confidentiality boundary. Chat and collaboration may recall across projects, while code sessions restrict project memories to their current project.

Maintain the conversation profile

Edit and save names or response preferences in the profile section. Change one value, then check the result in a simple conversation.

Profile language and interface language are different. By default, profile language is a fallback when a message is too short to identify its language. Always reply in this language overrides the current input language. Use Desktop settings to change menus and buttons.

To remove one value, use its clear action and review the confirmation. Deleting the profile clears that profile and offers an undo action after success. It does not delete chat history or other memory records.

If a change was not saved, resolve the save error. If refresh failed, the panel may still show the last synchronized profile rather than current values.

Curate existing memory records

Read the text, source, scope, and recall information before choosing an action.

Pin a small number of lasting facts, rather than temporary plans. When a fact changes, correct the specific entry. Inspect duplicate records expressing the old conclusion; editing one entry does not update all of them.

Old background must not override your explicit instructions in the current task.

Check recall and actual use

Open Recall test and enter a real question related to the saved fact:

Who is my product documentation written for?

Choose Test recall and inspect matching text, source, scope, and filtering explanations. The test merges eligible pinned entries and marks results in Memory list.

Distinguish three stages:

1. Saved: information exists in a profile or memory record.
2. Recalled: a query found candidates that survived curation into the test result.
3. Used: the actual task selected some information as context and generated its result.

Recall test is not a complete preview of the next answer. Relevance, negative feedback, and per-turn capacity also affect conversation selection. A match does not guarantee use. Recall counts and recent timestamps are clues, not proof that this answer followed the entry.

If the test reports filtering, inspect the reason. Hidden items, low-value records, unapproved AI replies, and evolution records can be excluded from user-memory results. Recording an AI reply does not make its unsupported claims facts about you.

After checking saving and recall, verify the deliverable in a real task. See the example in Long-term memory. Asking “Do you remember?” alone does not establish that later work will meet your requirements.

Hiding, clearing, and deletion

Different actions affect different objects:

• Hide a memory record: excludes local display and use; it does not permanently delete the backing record.
• Clear profile values: removes those values; deleting a profile does not clear all memory.
• Delete or archive chats: manages conversation history and does not automatically clean up memory.

For sensitive information, identify every object containing it and check the deletion scope offered for each. Local hiding, changing displayed text, or deleting a chat is not proof that all copies were erased.

Memory describes the active recall mode. Recall is local by default. EvoMap cloud search sends memories and query messages to EvoMap and uses credits; when unavailable, the page explains the local fallback. Local retrieval does not mean memory never leaves your device: information selected into task context is still sent to the model service used for that task.

Troubleshooting

Next: Profile and response style · Project working rules · Memory and identity