Mneme: The Shape of Memory in Machines

What memory means here

Mneme (pronounced NEE-mee) takes its name from the Ancient Greek mnēmē, meaning memory. An artificial intelligence model answers from the information available in its immediate working context. It does not automatically carry a complete record of earlier interactions from one answer to the next. A memory system reconnects the model with that past by deciding what to preserve, what to retrieve, and how to present it when the model needs it.

By shape, we mean the form that remembered information takes when it reaches the model. In this study, the model that answers is the reader, and the form in which stored history reaches it is the memory interface. Each experimental memory condition implements one memory method. The same past can arrive as a chronological log, a set of search results, linked facts, separate fields, or a compact account of the current state. Each form changes both what the model sees and what it must work out before it can answer.

The right fact may arrive too late, sit inside pages of routine history, or reach the model in a form that obscures how it relates to the question. This study follows the path from stored history to the model's answer and asks how much that final presentation changes the result.

The clearest difference appeared as history grew

Recent history gave the model a chronological window of the newest authorized events. At session 4, the generated history contained 9 event records. At session 64, it contained 194 raw records and 189 records after shared eligibility filtering. The longer window carried far more routine noise, and the 4,096-token condition could retain only part of it.

The difference was visible in both readers. From the 4-session to the 64-session checkpoint, recent-history accuracy fell from 71.88% to 50.74% for 4B and from 85.85% to 53.93% for 9B. The compiled working set did not show the same decline.

Compiled state used about one quarter as much input

Recent history used about 2,241 input tokens per question. Compiled state used about 575, roughly one quarter as many, while producing more correct answers. Input tokens are the pieces of text the model had to read before answering.

The compiled condition changed the serialization of the fused working set: it removed session, type, and provenance tags, added a working-set heading, and used compact bullets. This experiment measured their combined effect. A follow-up study will need to vary those elements one at a time to learn which parts supplied the gain.

What the result means

The memory interface changed what the reader had to work through. Recent history supplied a growing chronological window that included large amounts of routine noise. The fused conditions selected a much smaller set of records. Among those fused conditions, the compact compiled serialization outperformed the session-tagged hybrid form and the type-and-provenance-tagged form even though all three began with the same record ranking.

The result is therefore not evidence that one retrieval algorithm was better than another. It shows that the reader-facing serialization of the same fused selection coincided with large accuracy differences. Because the compiled treatment changed its heading, tags, and formatting together, the experiment cannot yet identify which element contributed most.

The reader mattered as well. The 9B model was more accurate than the 4B model under every deployable interface. Because the two models differ in more than parameter count, the study does not identify model size itself as the cause.

Limits and evidence

This study provides evidence about a memory-interface pattern, not a verdict on every possible memory system. It used two models from the Qwen 3.5 family and 256 lexical variations of one controlled history-and-question template. The confidence intervals describe variation across those template instantiations, not across structurally different kinds of history. Real conversations and projects will be noisier, more ambiguous, and less complete.

The scope-forgetting questions also did not distinguish the interfaces: every condition, including no memory, answered them correctly because the shared eligibility layer had already removed private and deleted records. That one uniformly correct family accounts for the shared 11.11% no-memory floor; it is not an estimate of chance performance. The result checks the common filtering rule, not the relative quality of the nine interfaces.

The study recorded 53 invalid structured outputs: responses that did not follow the required answer format closely enough for the automatic scorer to interpret them. The primary analysis counted them as incorrect. Removing them in a separate sensitivity analysis changed no paired estimate by more than 0.117 percentage points, so they did not materially affect the result.

We preserved the records behind the reported comparisons, uncertainty estimates, invalid-output treatment, and clean analysis rerun so that each result can be inspected.

The study design, estimands, bootstrap plan, and reporting specification were frozen before discovery, but the executable analysis was implemented after the run. We therefore describe this as design-frozen discovery, not an externally preregistered confirmation. A separate mixed-effects model failed to converge and does not support an additional result.

After the first analysis was complete, we copied the sealed evidence into a new environment on a different operating system and processor architecture. The complete analysis ran again from the original trials, code, seeds, and declared bootstrap procedure.

The rerun checks the path from preserved trial evidence to the reported results. It is not an independent reproduction, and it did not repeat the full GPU inference workload.

Sealed files checked

2,760

Hash mismatches

0

Tables reproduced exactly

10

Bootstrap arrays

Exact match

Where the research goes next

The experiment narrows the next set of questions. We now need to determine which parts of the compiled serialization produced the gain, whether the pattern survives structurally different histories, and whether different readers and questions need different forms of memory.