Memory is not a context strategy

An 8B-parameter model scores 42.8% on GAIA, a hard agent benchmark, while reading 6–7 times fewer input tokens than a standard ReAct agent. It does it by throwing its own context away.

That is one of four results, published within weeks of each other, that make the same uncomfortable point. More machinery around an agent's context does not reliably beat simply retrieving the evidence. Context has to be selected and pruned, not accumulated. Memory is a component. It is not a strategy.

Retrieval over raw history keeps winning

The paper with the sharpest name is Memory Tax: Dissecting the Net Value of Agent Memory. Its authors compared dedicated agent memory systems against naive retrieval over the raw history, across several benchmarks and model backbones. Their finding: many dedicated memory systems underperform naive RAG in many cases.

They split the effect in two. The memory benefit is what a system gains by changing how raw information is organised and accessed, for instance by distilling a preference. The memory tax is what it loses when that processing removes or obscures evidence the task later needs.

The point is the baseline. A system that restructures its inputs is making a bet that the restructuring keeps more than it throws away. Sometimes it does not.

An honest edge: the abstract page gives no magnitudes, and we have not read the full text. We can say the direction. We cannot say how big the tax is.

Accumulation decays

A second paper, Hippo, looks at the other common instinct: store every experience and let the agent draw on it. The authors report that this plain accumulation helps at first, then eventually falls below an agent with no memory at all.

Think of a drawer that is only ever added to.

Hippo's answer is consolidation. It compares new experience with what is already stored, merges overlaps, resolves conflicts, and attaches explicit conditions saying when a piece of guidance applies. The result is substantially fewer retained experiences than the accumulation baseline.

Note what that is. It is not more memory. It is memory that is owned, scoped and revised, which is closer to data governance than to storage.

The strongest result: clear the context

The third paper goes furthest. CORAL, accepted at COLM 2026, starts from the diagnosis that long-horizon agents fail from cluttered working memory. Its fix is a set of tools the agent calls itself: checkpoint the verified facts and progress, then reset the context and carry on from a clean state.

The reported numbers are the punchline. An 8B model reaches 42.8% average accuracy on GAIA using 6–7 times fewer input tokens than a ReAct agent. A strong proprietary agent gains 8.0 points on BrowseComp. The gains are largest on the hardest GAIA tasks.

The agent got better by forgetting on purpose. It kept what it had verified and dropped the rest.

A fourth paper points the same way from the cost side. Why Pay for Long Reasoning When Rewrite Do Trick? tests rewriting an agent's earlier reasoning into compact state records. For the tested models, compact histories usually shorten later reasoning. In software-agent deployments, compaction can reduce total spending, though task quality and total time, including the rewrite step, depend on the target and the compressor. Compaction is a tool with a failure mode, not a free saving.

What this means for a team building agents

Put the four together and the common thread is not "memory is bad". It is that what an agent sees is a decision, and every layer that makes that decision for you can make it badly.

For practitioners that comes down to a few habits:

• Benchmark against raw retrieval first. If a memory layer cannot beat plain retrieval over the original record on your task, it is adding a tax, not a benefit.
• Give stored knowledge conditions and an owner. A lesson without a statement of when it applies will be applied everywhere.
• Plan for deletion. Consolidation, expiry and resets belong in the design from the start, not in a clean-up project later.
• Treat compaction as lossy. Measure answer quality, not just tokens saved, and expect the result to differ by model.

None of this is exotic. It is how we would treat any other data feeding a decision: selected, scoped to the role, kept current, owned and versioned. Our view is that context management pays back before a bigger model does. These papers do not prove that, but they show how it goes wrong when context is treated as a pile that only grows.

Where the evidence stops

These are research benchmarks, not enterprise workloads. Two of the four papers give no figures on the pages we read, and we have read abstracts, not full texts, so we quote only the numbers the abstracts state.

A paper showing that some memory systems lose to naive retrieval says nothing about whether yours does. That is a claim to test on your own data, against the plainest baseline you can build.

If your agent's memory has only ever been added to, assume it is already costing you something. The first job is not a better memory. It is deciding what the agent is allowed to forget.

Kz