The underrated interface in agent memory is not capture. Capture is easy.
The hard interface is promotion: the moment a raw fragment becomes something the system is allowed to reuse.
Most personal knowledge systems avoid this step because it creates friction. That is exactly why it matters. Without a promotion step, the agent cannot tell the difference between a passing thought, a working draft, a sourced claim, a durable preference, and a decision that should affect future work.
Job
Promotion turns raw intake into reusable memory.
It should answer a narrow question: what can this system safely treat as context later?
Inputs
The input is a raw intake record plus any nearby context needed to judge it: original source, related notes, project state, date, and visible contradictions.
Processing
A promoted memory record should include:
- the claim, preference, pattern, decision, or project state being stored
- the source trail
- the date it became valid
- the date it should be reviewed again, if it can go stale
- the project or domain it applies to
- whether it is private, shared, or publishable
- confidence level
- contradictions or competing notes
- the reason it is worth reusing
This is where summarization has to be careful. A summary is not enough. The system needs to know what kind of memory it is creating.
"Mary prefers short paragraphs" is a preference.
"The team decided to use one orchestrator" is a decision.
"This source says retrieval uses semantic similarity" is a sourced claim.
Those should not live in the same shape.
Output
The output is a memory record with status.
Status matters. A useful system can distinguish active, draft, superseded, archived, private, and needs-review memory. Without status, old context keeps pretending to be current context.
Review Question
The review question is: would I be comfortable if an agent used this later without asking me to re-read the original?
If the answer is no, it may still belong in the archive. It does not yet belong in memory.
Failure Mode
The failure mode is mistaking organization for trust.
Tags, folders, embeddings, and summaries can make a messy archive feel intelligent. They do not make it reliable. Reliability comes from knowing which records have been promoted, why they were promoted, and when they should stop being used.