I keep seeing the same failure in every multi-agent setup I touch. Memory looks fine on day one. By week three it is half stale facts, half private context that should not have been written publicly, and half decisions that were superseded but never overwritten. Retrieval gets noisier. Users keep repeating context because the right fact ended up in the wrong scope. The recursion limit is not the problem here. The memory store itself is the problem.
The thing I changed that helped most was the simplest possible rule. Worker agents are not allowed to write to durable memory. They emit a structured memory event with a proposed scope and evidence, and a separate Memory Curator agent decides whether to write it, where to write it, or to discard it.
The four scopes I route into are agent repo memory (durable design rules for one agent), agent team memory (cross-agent procedures, handoff standards, safety rules), project memory (current state, decisions, risks for one engagement), and session scratch (temporary observations that probably should not survive). The mapping I had in mind was to organizational and human memory categories: individual specialist memory, transactive team memory (Ren and Argote), project memory, and short-term working memory.
The routing rule is conservative on purpose. If an event is temporary, unsupported, ambiguous, or contains private context, it goes to session scratch or gets discarded outright. Durable memory has to be earned. The schema is JSON with tagged fields for fact, decision, preference, risk, procedure, and hypothesis, plus an evidence reference and a proposed scope that the curator can override.
The reason I think this is the right architectural shape is that "what should be remembered, where, and for how long" is a different cognitive task from "do the work." When the same agent does both, the work agent biases toward remembering everything it produced. A dedicated curator whose only job is memory governance ends up much more conservative, and the store stays useful longer.
[link] [comments]