Most agent memory systems are built around the wrong goal: remember everything. Organizational memory does not work that way. It is a fact store, and the two operations that matter most are the ones almost nobody builds: rot and supersession.
Facts have wildly different lifespans. "We need more sales" is true today and stale in a quarter. A month-to-date sales report is fully invalidated the moment the next one lands. But yesterday's daily sales number is true forever. If your memory layer treats all three the same way, your agents either drown in stale context or burn tokens re-deriving the same facts in every session. Both failure modes look like a dumb agent. Neither is a model problem.
Supersession is not deletion. One closed deal can validate a hundred facts about your pipeline, your positioning, your pricing. If that deal falls through, the old facts come back. So you never delete. You mark a fact as superseded and keep the reversal cheap. History stays searchable underneath; only the working set stays lean.
There is an old format that got this right. Before dashboards, overnight batch jobs printed management reports on dot-matrix printers, and those stacks of paper sat on every manager's desk. That format is nearly perfect LLM context: dense, current, and explicitly superseding. The daily brief replaces yesterday's brief. The raw logs still exist if you need to drill in, but the thing you load into context is the fresh distillation, not the archive. Generating that report is now an agent workflow, and consuming it is what every other agent in the mesh does first thing in the morning.
I have argued before that general agent memory is unsolved while workflow-scoped learning is not, and this is the organizational version of the same point. You do not need a system that remembers everything about your company. You need a fact store where lifespan is a first-class property, where reports supersede reports, and where organizational context accumulates instead of getting re-derived. Build the rot algorithm before you build the bigger index.
Key takeaways
- Facts have different lifespans - a month-to-date report is invalidated by the next one, while yesterday's daily figure is true forever.
- Never delete superseded facts, because a reversed decision or a lost deal can make old facts current again overnight.
- The pre-dashboard management report, dense and current and superseding, is close to the ideal context format for an LLM.
FAQ
Why not just store everything an agent sees and let retrieval sort it out?
Because stale facts are actively harmful, not just noisy. An agent acting on last quarter's priorities makes confident, wrong decisions. Memory needs explicit rot and supersession semantics, not just a bigger vector index.
What is the difference between deleting a fact and superseding it?
Deletion loses history. Supersession marks a fact as replaced while keeping it retrievable, so when a deal falls through or a decision reverses, the older facts can be reinstated cheaply instead of re-derived from scratch.
Related Essays
Agent Memory Is Unsolved. Workflow-Scoped Learning Is Not
General-purpose agent memory is still a research problem. The opportunity is workflow-scoped learning that compounds — pick the constraint, and the memory problem stops being intractable.
Organizational Context Is the Hardest Problem Nobody Has Solved
Context management is the layer most consistently built in-house and least well served by vendors. It is not a search problem. It is a knowledge management problem.
Stop Re-Deriving the Same Facts
The practical win in agent memory is not total recall. It is a shared fact store with a lifecycle, so agents stop re-deriving what the organization already knows.
Drowning in pull requests that need your review? Try Tembo Review, a beautiful AI-assisted PR review tool unlike anything you’ve used.