TARL: Transaction-Aware Reliable Ledgers for Executable Memory Management in Long-Term Agents
TARL is a memory state update framework that maps each statement to one of five executable actions: add, ignore, revise, reject, or defer. It identifies affected memory, resolves temporal scope, compares source reliability, and updates accepted, pending, and rejected ledgers. TARL is trained by comparing memory states from alternative update operations. TARL-Mem is a benchmark with fine-grained action labels and next-state targets. Evaluations show TARL improves action prediction, state recovery, reduces memory pollution, and preserves conflicting information.
Persistent memory helps long-term agents retain knowledge, but a single update error can distort future retrieval and reasoning. Most systems reduce memory updating to a binary Write/Hold decision, failing to distinguish whether new information should be added, ignored, used to revise an outdated belief, rejected as unreliable, or deferred for verification. TARL introduces a framework that maps each statement to one of five executable actions, managing accepted, pending, and rejected ledgers. It is trained by comparing memory states from alternative operations. TARL-Mem provides a benchmark with fine-grained labels. Evaluations across in-domain, cross-source, temporal, counterfactual, and sequential settings show improved action prediction, state recovery, reduced memory pollution, and preserved conflicting information.
TARL replaces binary memory update decisions with a five-action taxonomy (add, ignore, revise, reject, defer) and uses ledger-based state tracking. Training involves comparing memory states resulting from different update operations, encouraging the model to select the operation that leads to the correct next state. This approach addresses the ambiguity of binary labels and enables more precise memory management.
Current long-term agent systems suffer from memory corruption due to simplistic update mechanisms. TARL's fine-grained, transaction-aware approach could improve reliability in applications requiring persistent knowledge, such as personal assistants, customer support bots, and research tools. The introduction of a dedicated benchmark (TARL-Mem) may drive standardization in memory management evaluation.
Improved memory reliability can reduce errors in long-running AI agents, lowering operational risk and increasing user trust. This is particularly valuable for enterprise agents handling sensitive or evolving information, where memory pollution can lead to costly mistakes.
Observable next signals include adoption of TARL-like memory management in agent frameworks, integration with retrieval-augmented generation (RAG) systems, and further research on memory consistency in multi-agent settings. The benchmark may be used to compare commercial and open-source agent memory solutions.