EVO CAPABILITIES / Short-term context and long-term memory
Remember the context worth keeping.
Long-term memory is most useful when you can understand it, correct it and choose what should influence future work.
Mac development · See current scopeThe real problem
Repeating your audience, preferred style or project decisions makes every new request feel like a first meeting. Keeping everything creates a different problem: a temporary assumption can follow you long after it stops being true. A useful AI memory must preserve important context while allowing it to change, rather than turn every sentence into a permanent fact about you.
What chat history leaves you to do
Recent conversation helps an assistant follow the immediate exchange. It is not the same as a reviewed preference that should carry into another task. A large transcript also mixes questions, rejected ideas, quoted material and final decisions. Without that distinction, retrieving more text can repeat an old mistake instead of making the next answer better.
How EVO approaches it
Separate today's conversation from lasting context
EVO combines a limited recent exchange and saved Work state with selected, approved memory. A draft assumption can stay with the current task; a confirmed preference can support later work. Reference documents live separately, so importing a document does not automatically turn its contents into facts about you or include them in every conversation.
Retrieve a useful subset
Supported recall selects a small set of relevant records using the current task's cues and available scope. It checks whether a record is active, appropriate to use and still valid before including it. This makes the selected context inspectable. It does not promise a perfect understanding of every past conversation or unlimited recall of your entire archive.
Treat correction as part of memory
You can inspect memory and ask EVO to add, correct or forget an item. The request is prepared for review; its appearance in the composer does not save a change. The memory decision remains separate from the conversation. Expired, suppressed or unconfirmed material should not quietly re-enter the supported recall flow as a confirmed fact.
Share understanding, not extra authority
Approved memory can give specialist conversations a common starting point: an audience, a decision or a writing preference. A researcher and writer can then work from compatible context. Remembering that you prefer decisive answers never grants permission to open another file, send a message or share private material with a cloud service.
A concrete example
ILLUSTRATIVE WORKFLOW · NOT A CUSTOMER RESULT
Illustrative workflow: your audience changes
You previously focused on individual creators. The next product launch is aimed at small business teams, but an older preference still influences drafts.
Open Memory and inspect the relevant saved understanding rather than guessing which old chat caused it.
Ask EVO to replace the audience preference and review the proposed change.
Use the confirmed preference in a later writing task, while keeping temporary campaign details with that Work.
What you take away
The useful result is an explicit, reviewable change to future context. You can see what should now guide a draft, instead of repeatedly correcting the same background in unrelated conversations.
Your choices
- Read saved understanding and request additions, corrections, suppression or forgetting through the supported review flow.
- Keep reference files separate and choose whether a task may use their content.
- Review restored memory before it is reused; a backup should not silently reactivate something you had chosen to forget.
Current scope
Implemented foundation
- A real local-model test created a preference proposal while leaving confirmed memory unchanged before approval.
- Correction, deletion and account isolation have controlled tests. The complete App journey has not yet been qualified for public release.
Next milestones
- Public access and broad long-history retrieval remain unfinished. Relevance and answer quality still need customer evaluation across varied real tasks.
- Forgetting changes later use of the selected memory; it does not erase original conversations, separately saved documents, exports or older backups. Those have their own deletion controls.
These are development capabilities, not a public release. See platform availability before requesting a trial.
Platforms and availabilityRelated questions
Can EVO remember useful context without loading every conversation?
Yes. EVO combines a limited recent conversation with saved Work context and approved memories, so durable preferences can remain useful beyond one exchange.
Only selected, relevant material within the current context limit is included. Saved memory and the complete conversation archive serve different purposes.
An approved preference to compare primary sources can inform a new research brief without replaying the conversation where you first expressed it.
Review saved items and ask EVO to correct or stop using them. Restored memories need confirmation again.
This is limited recall, not automatic understanding or retrieval of every old chat. A missing item may need to be selected or supplied again.
How does EVO find relevant context in a long history?
EVO uses limited recent conversation, the selected Work and approved memory selected for the task, rather than sending the entire archive on every request.
Supported search looks within material already available to the current task. A reviewed continuation can add relevant prior results with their sources.
To revisit a requirement from an older project, reopen that Work or provide the relevant excerpt instead of expecting every old message to be present.
Select the project, correct a missing reference and review what context is being used.
Universal semantic search across all conversations is not a qualified capability. Relevance can be missed, and a long saved history does not mean unlimited model context.
What is the difference between short-term context and long-term memory?
Short-term context helps with the current exchange. Approved longer-term memory preserves selected facts or preferences worth using beyond that exchange.
Recent conversation, saved Work state, reference files and memory are separate sources with different selection and retention rules.
Today’s temporary deadline belongs to the task; your confirmed preference for concise summaries may be useful in later work.
Review what is saved as memory and ask for corrections, suppression or deletion instead of assuming every conversation becomes a permanent preference.
Recall has limits. A provisional guess is not a confirmed fact, and a saved item is not guaranteed to appear in every response.
How do I correct or delete what EVO remembers?
Open Memory, inspect the item and prepare a correction or forgetting request in conversation. Review the proposed change before it becomes active.
Preparing the request keeps it as an unsent draft and preserves existing unsent text. It does not save a memory or send automatically. Suppression stops future recall; deletion removes the owned memory content while preserving relevant audit metadata.
If your preferred project language changes, replace that preference instead of asking every specialist to ignore an obsolete one.
You choose the item and the intended change. Files, source conversations and backups can be managed separately.
Forgetting a memory does not erase the original conversation, external copies or older backups. It is not a promise of forensic erasure.
Do EVO’s specialists share my memories?
Specialist views are internal experiments, not a first-release promise. They cannot turn an inferred preference into approved memory or grant a new permission.
The current focus is the single conversation, its chosen context and reviewed memories.
Try the workflow
Bring a task you want to move forward.
Explore the Mac release, or tell us which workflow you would like to evaluate. An application does not guarantee an invitation.