Belief memory, hierarchical memory, and the case for giving agents control over their own context
One paper models agent memory under partial observability as belief rather than record. Another builds a hierarchical multi-agent memory for long-context reasoning. A third argues plainly that agents need memory control over more context than they are currently given.
Belief memory is the sharpest of the three ideas. An agent operating with incomplete information does not have facts to store, it has estimates with uncertainty attached, and a memory system that records them as facts destroys the uncertainty that made them useful.
The hierarchical approach attacks the same problem structurally — different retention policies at different levels, so that detail can be discarded without discarding the summary that made the detail findable.
And the control argument is the one with the clearest engineering implication: an agent that cannot decide what to retain is at the mercy of whatever heuristic the harness applies, which is why memory-as-cognition rather than memory-as-tool keeps recurring across this literature.
arXiv — Belief memory: agent memory under partial observability → · arXiv — HMARS: a hierarchical multi-agent memory system for long-context reasoning → · arXiv — AI agents need memory control over more context →