An AI memory should tell you where it came from
Remembering a decision is only part of the job. You also need to know who said it, what it meant, and whether it still applies.
A remembered sentence can leave out the important part
Imagine asking an agent when a project should launch. It retrieves a memory saying, “The launch is scheduled for Friday.” That sounds useful. But was Friday a commitment, a suggestion, or a target conditional on passing a review?
A short memory is a claim about the work. It may be written directly by an agent or extracted from a longer source. In either case, it can lose context. The ability to retrieve it later does not establish that it is accurate.
For decisions with consequences, we want a route back to the evidence. In Space, a claim derived from a saved document can point to the specific revision used to create it. Directly written claims can have different evidence; a document source is not guaranteed for every memory.
Follow the memory to the saved revision
Consider this illustrative sequence:
- Revision 1 of a launch plan says: aim for Friday if the access review passes.
- A memory is extracted from revision 1 and shortened to: launch Friday.
- Revision 2 moves the target to Monday because the review is still open.
- A later agent session retrieves the original memory while answering a launch question.
A source link preserves the trail, not the truth of a claim
- Original sourceFriday, if review passesRevision 1 contains a condition attached to the target date.
- Extracted claimLaunch FridayThe shortened memory loses the condition. Its source is still revision 1.
- Compare revisionsMonday in revision 2The newer plan moves the target because the review is still open.
Inspect the original evidence and the newer revision, then correct the remembered claim explicitly.
Opening revision 1 reveals the condition that the extraction missed. Comparing it with revision 2 reveals that the plan changed. A link to only the latest document would hide what the memory was originally based on.
Space retains document revisions. A source update can be shown alongside a memory, but it does not automatically make every earlier claim correct or current. Someone still needs to interpret the change and update or remove the stale claim.
Ask four questions before relying on a memory
These questions are useful whether you inspect the result yourself or ask an agent to do it:
- Origin: was this written directly, extracted from a source, or inferred in the current conversation?
- Evidence: can I open the source and identify the revision behind the claim?
- Freshness: has the source or the underlying situation changed?
- Scope: was the statement about this project and decision, or a different context?
If there is no source, the answer should say so. If the source is ambiguous, the agent should preserve that ambiguity. Confidence in a sentence is not a replacement for evidence.
Finding a memory and checking it are separate steps
Space searches eligible memory in the authorized workspace using meaning and matching cues, then returns a bounded set of results. A relevant result helps an agent find context. It can still contain an incorrect extraction or a decision that has expired.
That is why we want the source to remain close to the answer. “The old plan targeted Friday, conditional on review; the newer revision says Monday” gives a person more to work with than a confident answer based on the first match.
The example here illustrates a review workflow. It is not a claim that every agent already performs that comparison automatically, or that Space has measured perfect retrieval.
Make correction part of keeping memory
In the development app, you can inspect a memory, correct it, or remove it from active recall. Accepted memories do not wait in a human approval queue. That makes the ability to inspect and revise them afterward part of the everyday workflow.
Removing a memory from recall is separate from deleting the document that supplied it, its retained revisions, or backup copies. Those have their own retention and deletion behavior. Revoking an agent also cannot take back content it has already received.
Space is in development, and public sign-ups are not open. The principle guiding the work is simple: a person should be able to understand why something was remembered and change it when it stops being useful.