Agent memory needs facts, procedures and evidence

A preference, a procedure and the conversation behind a decision enter an agent's workflow at different points. A preference guides the current response; a procedure supplies steps; the original conversation supports a claim when someone disputes it. I would give each a separate home in persistent memory, with an explicit hand-off from stored material to the task that needs it.

The Hermes memory documentation provides a concrete example. It describes curated cross-session memory and access to earlier conversations. Its separate skills documentation covers procedural knowledge. This article examines those documented mechanisms; it is not a release announcement or an account of a deployment we tested.

Keep durable facts close at hand

Hermes separates environment notes, conventions and learned information from user preferences and expectations. The documentation assigns these roles to MEMORY.md and USER.md. A snapshot is supplied when a session begins, giving the assistant useful guidance without requiring it to retrieve every conversation that produced that guidance.

The snapshot matters because stored information and active context are not interchangeable. A fact can exist on disk without rewriting information already presented to a running session. When investigating behavior, checking the stored note is only part of the question. You also need to know what guidance the particular session actually received.

This memory is bounded. The documentation does not promise automatic compaction when it fills. Instead, the agent must consolidate or remove material before retrying an update. That makes curation part of the design: durable context should contain information valuable across future interactions, rather than an indiscriminate transcript of everything the assistant has encountered.

Give procedures a different home

A skill explains how to perform a category of work. Hermes organizes skills around SKILL.md, with supporting material such as references, examples and scripts where needed. The system loads procedural knowledge on demand. That is a different job from keeping a short user preference available throughout unrelated tasks.

Take a commercial proposal. A preference for direct language describes the user’s expectations. A checklist for checking the recipient, terms and attachments describes a procedure. A conversation in which a customer negotiated an exception is evidence about a particular agreement. Folding all of those into a general memory note risks turning a local exception into a standing rule.

The skills system’s progressive disclosure provides a useful organizational model. Keep broadly applicable guidance in persistent memory and retrieve a method when the task calls for it. This is our practical interpretation of the documented separation, not a claim that automatic skill selection will always identify the right procedure or apply it without error.

Retrieve what was actually said

The memory page documents session search backed by SQLite and full-text search. It describes results as actual messages returned from the database, without model summarization or truncation. That offers a different kind of support from a condensed memory note. The note may point to a decision; the original exchange helps establish its wording and scope.

There is no reason to invent an embedding pipeline or vector database to explain this capability. The documentation identifies the mechanism it uses. When choosing an agent, starting with a fashionable storage architecture can distract from the practical requirement: can the assistant recover the relevant evidence when a remembered instruction is challenged?

Recovered messages still need interpretation. An instruction might apply only to a particular project, customer or temporary circumstance. A historical message does not automatically override the present request. Retrieval makes the evidence inspectable; it does not grant every old statement permanent authority over future work.

Evaluate corrections as well as recall

A practical adoption exercise could introduce a clear preference, a short procedure and a decision recorded in a conversation. In later sessions, ask the assistant to apply the preference where relevant, retrieve the procedure for the appropriate task and locate the exchange when asked to explain the decision. These are suggested checks, not results we obtained.

Then correct an instruction and examine how the correction travels. The previous wording may still exist in an old session or a skill. Updating memory should not be confused with erasing history, and finding history should not justify restoring a cancelled rule. Useful persistence preserves the distinction between a record of the past and guidance that remains in force.

When a stored preference conflicts with a procedure, the agent needs a route back to the messages that established them. Curated facts, reusable skills and session records make that route inspectable without loading everything into everyday context. The remaining operational choice is what the agent may apply immediately and what it must recover from the original conversation before acting.