A preprint describes an AI memory system designed to keep the history behind a fact, including changes and conflicting observations, rather than collapsing that history into a single overwritten state. Called PolyMemDB, it is tailored for managing agent memory. Its central record is a graph, a network of entities and relationships, whose links retain a time-ordered sequence of observations. That design gives the system a visible trail of what was recorded over time when it handles a question.
The work is a methods and system demonstration built around three illustrative cases. One is a 48-session LongMemEval conversation. Another is an extended LongMemEval case about a graduation-trip itinerary. The third is a simplified 10-month history used to ask whether Alice enjoyed running. The cases show the software's intended workflow, but the displayed graph, retrieval output, itinerary view and reliability scores are not independently validated. This is evidence that the mechanisms and outputs are presented, not a measured verdict on accuracy or performance against other systems.
The record behind a memory
PolyMemDB's defining choice is a polyglot storage layer, meaning that it routes different forms of memory to different database systems. Graph data goes to Neo4j, probabilistic data to ProvSQL, spatio-temporal data, combining place and time, to MobilityDB, vector data to ChromaDB, and raw context to object storage. The architecture is meant to support heterogeneous long-term memory management while keeping those data types in the stores assigned to them.
Before storage, an ingestion pipeline uses large language models and vision-language models to turn incoming material into structured records. The method includes aligning entities that refer to the same thing, resolving coreference, recovering omitted information, extracting named entities, normalizing facts and estimating confidence. In practical terms, it prepares the raw conversation for later retrieval and evidence scoring by deciding how facts are represented and how much confidence is attached to them.
For questions, the response path uses a top-down, three-tier federated retrieval mechanism. In plain language, it searches through successive retrieval layers rather than treating every stored item as the same kind of memory. In the first scenario, the output also parses the reasoning chain, showing how the language model arrived at its final answer. The paper presents that chain and its supporting records as part of a provenance-preserving response path.
Three ways the system is shown at work
In Scenario 1, the 48-session conversation produces a reported memory graph containing 229 entities and 221 relationships after ingestion. A sample retrieval output is given as 20%, alongside the reasoning-chain display. The analysis does not independently validate the answer's correctness, nor does it validate the graph's accuracy or completeness. The numbers therefore describe the scale and presentation of this demonstration, not a verified accuracy rate or a result that can be generalized to other workloads.
The second demonstration turns the graduation-trip case into a planning view. PolyMemDB displays a structured itinerary and an interactive dashboard; when the planning period is filtered, 5 Italian cities are highlighted and a fine-grained fact list is tied to the underlying evidence. It is an example of spatio-temporal filtering, but the supplied analysis reports no formal accuracy or usability evaluation for the display. The interface is shown on the case, not assessed as a reliable tool for users.
The third demonstration focuses on disagreement inside the memory record. It asks whether Alice enjoyed running across a 10-month history and uses a temporal-decay factor of lambda = 0.8, semiring aggregation and a four-dimensional cognitive distribution. Temporal decay calculates dynamic weights from a fact's initial confidence and its distance in time. The engine then aggregates positive and negative evidence into four intermediate cognitive states and a Net Evidence Reliability score.
The illustrative output reports 47.2% positive drive, 52.3% evidence conflict and net reliability of approximately 0.22. Those values are internal scores from the example. The analysis reports no statistical uncertainty or external validation, and it does not establish that a reliability value of about 0.22 predicts real-world truth. A system can display a nuanced confidence picture without the demonstration showing that the picture is correct.
A design demonstration, not a verdict
Taken together, the cases show the machinery PolyMemDB sets out to build: retained memory history, specialized storage, evidence-linked displays and conflict scoring. They do not show improved factual accuracy, lower hallucination rates, faster responses, lower token costs or better recall than another system, because no comparative evaluation is reported. The document is an arXiv version 1 preprint dated 26 Aug 2026 and includes a source-code URL. Questions about performance on larger workloads and sensitivity to its decay and confidence settings remain open.
Paper data and sources
Original title: PolyMemDB: A Polyglot Database System for AI Memory Management
Authors: Yu Wang, Jiaheng Lu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text