Which kinds of memory you're actually keeping

Working context, durable facts, decisions, preferences, events and archives all live for different lengths of time and get retrieved in different ways. I'd sort those categories out before picking a database or an embedding model, and then decide which ones can share storage without also sharing retention or retrieval rules.

Working context belongs to the task in front of the agent: the immediate goal, the open plan, recent tool results and any temporary assumptions. When the task ends, it should be compacted or cleared. If you promote all of it into durable memory you end up keeping abandoned approaches and a lot of routine execution noise.

Durable factual records describe things and their current properties. A person's approved contact preference, a project's status, where a system's repository lives, who owns a service. Each one needs a source and a way to correct it. Some only hold for a while, too. A current owner might be valid from one date until someone else is assigned.

Decisions deserve their own record type. I'd keep the choice, its scope, the date, who made it (or under what authority), the reasoning where you have it and the alternatives it affected. A decision can be reversed later and still matter historically, so you don't want it quietly overwritten.

Preferences need their scope written down. "Use short updates for this project" is a different thing from a permanent personal preference. Store who said it, which context it applies to and whether they said it outright or the system inferred it. Inferred preferences should carry less authority and be easy to remove.

Events are the record that something happened: a deployment, an approval, a handoff, a failed run. Keep them as chronology and audit trail. A controlled write can then use a deployment event to update a separate current production version record, rather than asking the model to work out the current version from a list of events.

The archive is where source documents, transcripts and older records sit for search and checking. It can get big, and that's fine, because the default retrieval path isn't pouring it into every session.