代理框架與記憶
TARL Splits Agent Memory Updates Into Five Transaction Types, Raising Next-State Accuracy to 66.21%
Rather than managing agent memory as a binary “write or hold” decision, TARL explicitly distinguishes among appending, ignoring, revising, rejecting conflicts, and deferring for verification. The authors’ experiments show that the memory contamination rate fell to 25.24%, but the model and dataset are currently available only through the paper’s supplementary materials.
When long-running agents write incorrect information into memory, the contamination can be repeatedly amplified during subsequent retrieval and decision-making. The newly proposed Transaction-Aware Reliable Ledgers (TARL) therefore replaces the conventional Write/Hold decision with five executable transactions: `append`, `noop`, `revise`, `reject_conflict`, and `defer_verify`. Two pieces of information might both be labeled “write,” but adding a new fact and replacing an outdated one require different state transitions. Likewise, “do not write” might mean that the content is duplicated, the source is untrustworthy, or the claim still requires verification.
TARL first identifies the existing memory slot related to a candidate statement. It then compares new and existing evidence using time ranges, sources, confidence, and conflict signals before routing the data to three ledgers: accepted, pending, and rejected/history. Content superseded by a new fact does not simply disappear; it is retained with its source in the history ledger. During training, the system also simulates the next state produced by each candidate transaction and uses the correct resulting ledger state to supervise selection, rather than merely requiring the action label to be classified correctly.
The accompanying TARL-Mem dataset derives fine-grained labels from HaluMem, LoCoMo, and LongMemEval. In the paper’s primary comparison across five independent training runs, the authors report a five-class Macro F1 of 0.8286 and full next-state accuracy of 0.6621, both outperforming seven memory baselines. The contamination rate for accepted memories was 0.2524, which still means that roughly one-quarter of predicted commitments should not have entered the trusted ledger. In a separate closed-loop evaluation comprising 14 trajectories with 200 consecutive transactions each, TARL achieved a Stateful Rollout Score of 0.7258.
From an engineering perspective, TARL’s value lies in turning “memory updates” into an auditable state machine, making it easier to determine whether an error occurred during writing, revision, or retrieval. However, the reported results depend on data transformations and an executor created by the authors, and the system has not yet been shown to handle malicious sources on the open web, compound statements, or large-scale concurrent updates. A complete standalone repository has also not yet been released. The next issues to watch are whether these transaction semantics can be integrated with existing vector databases, and how pending memories will be governed with respect to verification, expiration, and concurrent conflicts.