THE SIGNAL IN ONE SENTENCE
Replacing an agent’s model can change how well its old memory works, even when every stored file remains exactly where it was.
01
WHAT ACTUALLY CHANGED
Researchers tested whether memory created for one language model remains useful after an agent moves to another. The controlled study used 48 synthetic histories with randomized answer codes, which let the authors score retrieval and answering without relying on another model’s opinion.
They compared four approaches: feeding the complete raw history into the context window, retrieving stored chunks, compressing the history into natural-language notes, and writing facts into a fixed-schema knowledge graph. Two open models below 10 billion parameters took turns constructing and reading the memories.
The fixed-schema knowledge graph was nearly unchanged after a writer swap, moving by only 0.0004 plus or minus 0.0020 in accuracy. Natural-language notes behaved much less predictably. One migration direction improved by 9.91 percentage points, while the reverse direction declined by 13.28 points.
Embedding changes created another migration problem. A half-old, half-new index captured only a 4.96-point improvement, compared with 11.90 points after full re-embedding. The authors traced most of the notes deficit to information lost when the memory was constructed, while most of the retrieval deficit came from failing to retrieve information that was still present.
02
WHY THIS MATTERS
Agent memory is not merely a folder full of facts. The model, summarization method, embedding system, retrieval rules, and storage format all influence what the agent can recover. Change one component and the same archive can begin producing different answers.
That creates an invisible risk for teams that upgrade models frequently. The new model may score better on a benchmark while quietly becoming worse at reading months of accumulated project history. Nothing looks broken until the agent forgets the exact decision somebody assumed was durable.
The study suggests a practical hierarchy. Keep raw evidence whenever possible, rebuild model-dependent indexes during migrations, and store critical facts in explicit structures that leave less room for interpretation. The boring database habits are returning, wearing an agent-memory badge.
03
WHERE IT COULD HELP
- Regression-test memory before changing an agent’s model
- Preserve original histories alongside generated summaries
- Rebuild embeddings instead of mixing incompatible versions
- Store critical facts in explicit schemas with traceable sources
KEEP A HAND ON THE WHEEL
The experiment used synthetic histories and two relatively small open models. It does not prove that every production agent will fail this way. Teams should test their own data, retrieval stack, and migration direction, especially when old memories can trigger consequential actions.
04
TERMS WORTH KEEPING
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 7, 2026.
PUBLICATION RECEIPT: Revision 1. Approved by Zak and published September 7, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US