AuthorsYan Zhou, Yue Ouyang, Kaiyang Zheng, Suncheng Xiang
TL;DR
TEPA uses conflict-keyed lifecycle revocation to keep only valid memories active, reaching 0.950 success under full reversal vs 0.210 for append-only memory.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Append-only and last-write-wins memory fall to 0.210 success during full reversal, below the no-memory baseline of 0.309, revealing memory pollution.
In this regime, language agents retrieve stale evidence as prompt context, causing conflicting facts to dominate decisions and making persistent memory actively harmful.
HOW IT WORKS
TEPA represents observations as keyed precedents with explicit lifecycle state, separates an active set from a revoked archive, and uses a retriever plus trial-validated promotion to manage validity.
You can think of TEPA like a card catalog where each card tracks a fact’s key, success statistics, and status, and stale cards are moved to an audit drawer instead of the main shelf.
This conflict-keyed lifecycle mechanism lets TEPA falsify and revoke outdated facts, something a plain context window or append-only cache cannot do while still preserving history for audit.
DIAGRAM
This diagram shows how TEPA updates precedent states, increments support or conflict counts, and revokes stale precedents during each episode.
DIAGRAM
This diagram shows how TEPA is evaluated on controlled drift, executable drift, preference updates, and MemoryAgentBench with shared baselines.
PROCESS
Episode Stream and Context
TEPA receives an episode et with visible context ct and uses the retriever to pull relevant precedents from the active set At.
Policy Execution and Outcome
Conditioned on ct and retrieved precedents, TEPA’s policy πm selects action a(m)t and observes binary outcome y(m)t plus evidence xt.
Conflict Key Extraction
TEPA applies extractors κ and ν to xt to obtain conflict key kt and value vt, then updates keyed precedents with support or conflict counts.
Lifecycle Update and Revocation
Using the Beta-Bernoulli posterior q(p), TEPA performs lifecycle update, revoking precedents into the revoked archive Rt or promoting hypotheses to the active set.
KEY CONTRIBUTIONS
Formulating Memory Pollution and Pollution Index
TEPA defines memory pollution as a validity-state failure and introduces a phase-wise memory pollution index MPIm,ϕ, showing append-only success dropping to 0.210 vs 0.309 for no memory.
Conflict-Keyed Revocable Evidence Memory
TEPA proposes keyed precedents with explicit lifecycle states and revocation, separating an active set At from a revoked archive Rt for falsifiable long-term memory.
Lifecycle Evaluation Across Four Conflict Settings
TEPA is evaluated on controlled drift, executable drift, preference-update streams, and MemoryAgentBench SH6k, reaching 0.950 reversal success and 0.890 substring exact match on SH6k.
RESULTS
Full-reversal success
0.950
+0.740 over Last-write-wins
Executable reversal success
0.950
+0.747 over Append-only
Preference full-reversal success
0.872
+0.034 over No memory
MemoryAgentBench SH6k substring EM
0.890
+0.307 over Append-only
These metrics come from controlled hidden-regime drift, real file-backed executable drift, preference-update streams, and MemoryAgentBench SH6k, which test stale-conflict consolidation. The 0.950 reversal success and 0.890 substring exact match show TEPA keeps memory useful where append-only and last-write-wins become worse than no memory.
BENCHMARK
These metrics come from controlled hidden-regime drift, real file-backed executable drift, preference-update streams, and MemoryAgentBench SH6k, which test stale-conflict consolidation. The 0.950 reversal success and 0.890 substring exact match show TEPA keeps memory useful where append-only and last-write-wins become worse than no memory.
BENCHMARK
Success rate across controlled drift, executable drift, preference updates, and MemoryAgentBench SH6k.
KEY INSIGHT
Under full reversal, append-only and last-write-wins memory drop to 0.210 success, below the no-memory baseline of 0.309, yielding a pollution index of 0.318.
This is surprising because persistent memory is usually assumed to help agents, yet TEPA shows that without revocation, stored experience can be strictly worse than forgetting.
WHY IT MATTERS
TEPA unlocks long-term memory that can falsify, audit, and later re-promote evolving knowledge by tracking validity state per conflict key.
Builders can now design agents whose memories remain trustworthy under regime shifts and preference changes, instead of relying on brittle append-only caches or global resets.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…