Memory enables language agents to carry useful distinctions from past experience into future decisions. Its value depends on which distinctions survive, whether a bounded agent can use them, and what can be recovered or revised as demands evolve. We develop a decision-theoretic framework that follows this value from raw history through persistent state, operational use, and costly recovery and revision. Original and retained value are expressed in units of avoided task loss; risk-cost frontiers describe attainable recovery and preservation across uses. Unknown demand yields value profiles, while interactive feedback changes the distribution of future work. A cost-aware simulation condition and five minimal counterexamples expose the roles of information timing, retrieval access, component interaction, feedback and demand uncertainty. In particular, a pointer with zero mutual information about the answer can still reduce the cost of exact recovery. Diagnostics on LoCoMo, a two-project update task and longitudinal cases exhibit failures of preservation, use and measurement. At fixed retrieved evidence, a saved ledger changes subsequent answers; at fixed agent outputs, an evaluator revision changes scores. The framework connects these distinctions to operational risk bounds and testable principles for allocating memory, preserving recovery options and revising evidence dependencies.
Figures & tables
Figure 1: Memory value across information, operation and measurement. The operational system Σ includes writing, storage, recovery, revision and reading. Its actions affect the environment and later demands; J measures the resulting trajectories under a task loss and reference labels. The outer boundary records comparison conditions. The dashed query input denotes the query-aware alternative. The lower strip compares potential risks under matched assumptions; O denotes new revision evidence.
Environment
p0
p1
Policy 0 risk
Policy 1 risk
E+
0
1/4
0
3/4
E−
3/4
1
3/4
0
Table 1: Identical forced-replay behavior is compatible with opposite interactive rankings.
Figure 2: Four mechanisms shape memory value. (a) Query timing changes required state size. (b) A two-bit pointer with I(M;Y)=0 reduces risk under paid recovery. The horizontal axis is the maximum allowed reads k ; zero-error expected reads are 5/2 for search and 1 with the pointer. (c) Memory and reader improvements interact. (d) Feedback changes the ranking compatible with fixed replay. All values follow the constructions in Section 5 .
Component
Question resolved
Information and revelation order
What can the writer know before committing to memory?
Persistent state and side information
Which stored or externally supplied resource supports the answer?
Recovery and update interfaces
What can be reread, refreshed, written back or asked of the user?
Policies and computation
Which decisions are feasible under the available reader and tools?
Demand and environment
Which future uses and feedback processes enter the evaluation?
Consequences and loss
What makes an answer, action, delay or failed delivery costly?
Table 2: A comparison contract connects an evaluation result to its interpretation.
Benchmark
Useful observables
Value question it can help operationalize
LoCoMo
Multi-session history and annotated evidence
What does a saved state support, and what can additional retrieval recover?
LongMemEval
Updates, time-dependent facts and abstention
Does later use follow the valid state, and when is additional evidence needed?
MemOps
Operation targets, traces and resulting states
What consequence should a memory operation have, and which other facts should persist?
Table 3: Public benchmarks as sources of evidence for the framework. The behavioral pilot in this paper uses LoCoMo.
Mean bytes supplied
Evidence available to the reader
Ledger
Archive
F1 (0–1)
Abstentions
No history
0
0
0.000
12/12
Saved ledger
3,823
0
0.016
11/12
Ledger + bounded recovery
3,823
4,065
0.329
3/12
Complete history
0
97,355
0.123
10/12
Annotated support reference
0
759
0.333
5/12
Table 4: A fixed-reader LoCoMo diagnostic: 12 questions per condition. Mean UTF-8 byte counts are rounded to the nearest byte and exclude the question. Archive bytes measure supplied evidence. The first five rows share a reader prompt; the lower rows are diagnostic controls on the same questions.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Postcheck
Pairs
Fixtures
Inspection result
Over-forgetting
24
20
Retained information is available; the deletion phrase refers to another target or operation.
Disclosure
11
10
The gold answer repeats a forgotten detail; task guidance conflicts with the disclosure rule.
Appendix
Table 5: Deterministic postcheck flags under gold-answer replay. The denominator is 400 question pairs from 80 fixtures. Semantic classification concerns the flagged diagnostic.
Long-horizon language agents must operate under limited runtime memory, yet existing memory mechanisms often organize experience around descriptive criteria such as relevance, salience, or summary quality. For an agent, however, memory is valuable not because it faithfully describes the past, but because it preserves the distinctions between histories that must remain separated under a fixed budget to support good decisions. We cast this as a decision-centric rate-distortion problem, measuring memory quality by the loss in achievable decision quality induced by compression. This yields an exact forgetting boundary for what can be safely forgotten, and a memory-distortion frontier characterizing the optimal tradeoff between memory budget and decision quality. Motivated by this decision-centric view of memory, we propose DeMem, an online memory learner that refines its partition only when data certify that a shared state would induce decision conflict, and prove near-minimax regret guarantees. On both controlled synthetic diagnostics and long-horizon conversational benchmarks, DeMem yields consistent gains under the same runtime budget, supporting the principle that memory should preserve the distinctions that matter for decisions, not descriptions.
Mingxi Zou, Zhihan Guo, Langzhang Liang +6
Fudan University · The Chinese University of Hong Kong · Meta AI +3
Persistent language agents need stored experience to remain available across time, while each answer requires evidence suited to a particular question. A superseded fact can mislead a current-state answer and still be essential for a historical query. We present RD-Forget, a training-free framework that separates what an agent stores from what it uses. A retained source archive preserves observations, and a query-conditioned memory view controls their influence on the current answer. A frozen language-model curator extracts relevant evidence, groups facts into semantic slots, and preserves the relations needed for multi-hop reasoning. Same-slot replacement links suppress superseded values in current-state contexts, while intent-aware retrieval makes earlier evidence eligible again. A rate-distortion formulation guides construction of the answer-time view within a memory budget. Experiments span conversational memory, knowledge updating, fact consolidation, long-context reasoning, and personalization under a shared answering pipeline. The results associate accurate answers with both query-relevant evidence construction and control over obsolete alternatives. Configurations without forgetting or query conditioning have the largest score deficits, while slot grouping, historical access, and relation preservation contribute complementary functions. Retaining history while selectively controlling its use offers a practical way to accommodate changing facts and future questions.
When documents supporting an agent's derived facts are revoked or replaced, should it repair memory or re-read current evidence? We introduce an evidence-revision evaluation on medication- and problem-list tasks from public ICU records. Under revocation, replacement and control events, we compare full and source-filtered re-reading with caching, rebuilding and graph-local repair across two 7B models. Memory is supplied in full without retrieval, and costs include ingest, revision and every use. On short records, local repair uses 5-10× fewer revision tokens than rebuilding, yet every memory pipeline costs at least twice full re-reading in held-out conditions. In a small pre-specified development sweep, adding task-ineligible documents extended records to about 10,000 tokens; at that length, memory's mean cumulative cost fell below full re-reading's after 2-14 uses, partly through truncated extraction, while source-filtered re-reading remained cheapest. In the replacement study, none of the four primary confirmatory tests reached statistical significance. These results show why the cost of agent memory after evidence revision must be assessed against source-filtered re-reading over the full pipeline.
Wenhui Chu
University at Albany, State University of New York · Department of Cybersecurity College of Emergency Preparedness, Homeland Security and Cybersecurity University at Albany, State University of New York Albany, NY 12222, USA