The Cow That Remembered the Gate

Imagine a cow that walks into the same locked gate every morning. A new model would be nice, but the first useful upgrade is simpler: remember that the gate was locked yesterday, and remember where the open one was.

Many AI agent mistakes are failures of context, not failures of raw reasoning. An assistant may have solved the same problem in an earlier session but cannot access the correction, or it may retrieve the wrong instruction for a different machine, project, or user.

Memory is not one big notebook

A candidate external learning system can separate several kinds of knowledge:

• Environment: what services, hosts, tools, and permissions are currently observed, with timestamps.
• Project: durable decisions, constraints, and active tasks.
• Episodes: what an agent attempted, what tools actually returned, and what happened next.
• Corrections: explicit user changes that supersede older assumptions.
• Skills: reusable procedures that have been reviewed and approved for an appropriate context.

A vector index can retrieve similar experiences; a graph can preserve relationships such as this correction supersedes that instruction or this procedure applies only to this host. Neither should be confused with a factual guarantee.

The difference between remembering and learning

Storing an episode proves only that storage worked. Retrieving it proves only that retrieval worked. Putting it into an outgoing request proves only that the model was shown the information.

A stronger claim, behavior improved, requires measuring whether future tasks actually succeed more often, with fewer repeated mistakes or less unnecessary work. The model's own assertion that it learned something is a particularly weak metric.

We are exploring an approach where a controller selects relevant, scoped evidence before a task; the model uses that evidence during the task; and a separate intake process records verified outcomes and corrections afterward. Model weights need not change for this kind of in-context adaptation.

A test that can fail honestly

Choose repeatable tasks with known answers and known traps. Run a baseline without retrieved history, then rerun with a bounded context drawn from prior episodes and corrections. Hold the underlying model and tools constant where possible.

Measure task success, wrong-environment actions, stale-fact retrieval, correction compliance, and token/latency overhead. Also test contradictory memories and missing storage. A self-learning system must not invent an old success just because its database is unavailable.

The field note

The cow does not need to become a different species every morning. It needs a reliable map, a record of which gates worked, and a way to notice when the farmer moves a fence.

Research status: architecture and evaluation questions. Any performance improvement remains to be demonstrated in controlled tests.