AuthorsYongye Su, Wujiang Xu, Chaoji Zuo, Elisa Bertino
TL;DR
ChronoMem uses commit-on-write snapshots plus semantic NL→version mapping to boost rollback-consistent QA on MemoryAgentBench to 55.1% vs 35.5% for RAG-only (+19.6 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM agents typically use append-only or overwrite-only memory, with no principled way to inspect, version, or revert prior states.
This makes long-term agents brittle under corrections, concept drift, and memory poisoning, because later interactions lock in faulty updates and contaminate downstream behavior.
HOW IT WORKS
ChronoMem builds on LocalMemoryService, SQLite memory store, Version Index, and a Reranker module to snapshot global agent memory on every write and expose rollback APIs.
You can think of ChronoMem like git for an agent’s memory: commits are whole-memory snapshots, and natural-language “undo” requests are mapped to version IDs via a semantic index.
This architecture lets ChronoMem restore a prior global memory state and enforce counterfactual behavior, something a plain context window or vector store cannot guarantee after exposure.
DIAGRAM
This diagram shows how ChronoMem maps a user’s natural-language undo request to a concrete historical version and restores that snapshot.
DIAGRAM
This diagram shows how ChronoMem is evaluated under the post-exposure protocol on LoCoMo and MemoryAgentBench.
PROCESS
CommitOnWrite
ChronoMem uses CommitOnWrite to append events to the log, build a commit descriptor, materialize a snapshot, and advance HEAD via LocalMemoryService.
Version Indexing
ChronoMem constructs semantic commit descriptors and stores them in the Version Index, backed by SQLite FTS and a vector embedding index.
NL Version Mapping
ChronoMem runs hybrid lexical and dense retrieval over commit descriptors, fuses ranks with Reciprocal Rank Fusion, and applies the Reranker module to select a target version.
RollbackToVersion
ChronoMem calls RollbackToVersion to load the snapshot, restore state, set HEAD, truncate later versions, and synchronize any retrieval backend.
KEY CONTRIBUTIONS
Version Control Abstraction for Agent Memory
ChronoMem introduces global snapshots, version histories, and commit semantics inside ADK’s LocalMemoryService, wrapping a SQLite memory store and retrieval-backed index backend.
Semantic Global Rollback Mechanism
ChronoMem provides rollback_to_version_by_nl_query, mapping natural-language undo requests to whole-memory versions via hybrid retrieval, Reciprocal Rank Fusion, and a cross-encoder reranker.
Post-Exposure Rollback Evaluation Standard
ChronoMem augments LoCoMo and MemoryAgentBench with evolving memory states and rollback tasks, measuring rollback-consistent QA and summarization under a post-exposure protocol.
RESULTS
Recall@1 on MAB
33.4%
+9.1 points over Hybrid (RRF) version selection
Recall@5 on MAB
60.2%
+6.4 points over Hybrid (RRF) version selection
Scope@2 on MAB
58.0%
+17.5 points over Hybrid (RRF) version selection
QA Accuracy on MAB
55.1%
+19.6 points over RAG-only with Qwen2.5-7B
On MemoryAgentBench, which tests accurate retrieval and long-range understanding under incremental updates, ChronoMem’s semantic version control improves both version targeting and rollback-consistent QA. The 19.6-point QA gain over RAG-only shows that restoring whole-memory snapshots matters beyond better retrieval.
BENCHMARK
On MemoryAgentBench, which tests accurate retrieval and long-range understanding under incremental updates, ChronoMem’s semantic version control improves both version targeting and rollback-consistent QA. The 19.6-point QA gain over RAG-only shows that restoring whole-memory snapshots matters beyond better retrieval.
BENCHMARK
Task-specific QA accuracy after rollback on MemoryAgentBench using Qwen2.5-7B.
BENCHMARK
Recall@1 for different version selection strategies on MemoryAgentBench.
KEY INSIGHT
ChronoMem’s prompt-only rollback baselines achieve as little as 0.8% QA accuracy on MemoryAgentBench, despite explicit instructions to ignore later information.
This is surprising because many practitioners assume careful prompting can undo contamination, but ChronoMem shows that without architectural state restoration, instruction following barely helps.
WHY IT MATTERS
ChronoMem gives LLM agents a natural-language "undo" button over global memory, enabling counterfactual behavior after they have already seen future interactions.
Builders can now implement auditable, reversible long-term personalization and recover from memory poisoning or bad updates, instead of hoping prompts or ad-hoc deletions are enough.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…