Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteIn Haoning Kan’s comparison of six AI agent-memory configurations, the highest reported LoCoMo score was 90.75% for ListWise + Rewrite at a 0.8 threshold. But the score was effectively tied with a simpler baseline within the experiment’s reported run-to-run variation. The more useful lesson is about keeping memory up to date: a system can score well while retaining contradictions or deleting details that still matter.
What the six strategies actually changed
The comparison was about external long-term memory: a system extracts facts from conversations, stores them, and retrieves relevant entries later. The six configurations were not six different memory architectures. They varied along two axes: how a new memory is compared with existing ones, and what happens when the system decides that an old entry overlaps with it.
As an Amazon Associate I earn from qualifying purchases.
Judgment: PointWise or ListWise
PointWise compares a new memory with candidate old memories one at a time. ListWise judges a batch of candidates together, potentially identifying more than one entry that needs attention in a single judgment call. Neither approach can act on a relevant old memory that candidate retrieval fails to surface.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Update: Rewrite or Replace
Rewrite merges new information with the old entry, aiming to retain useful details while updating what has changed. Replace substitutes the new entry for the old one. Replacement is simpler, but it can discard information that the new entry does not repeat.
#1 Best Overall
Kan uses mem0 v2.0.7, MemOS 2.0.23, and signetai 0.123.22 as product examples of different handling: respectively, no standalone deduplication in the described open-source path, pair-by-pair PointWise judgment, and batch ListWise judgment with replacement. The six measured configurations, however, were reproduced on NeatMem to focus on strategy rather than product-specific prompts and retrieval implementations.
Reported results: the score gap was small, the call counts were not
Kan’s DEV article is marked “Sep 26” without an explicit year; examples and comments refer to 2026. The figures below are author-reported means across five LoCoMo runs, not an independent replication. Every configuration used 1,400 extraction calls. The experiment used identical write, embedding, and scoring models, began each run with an empty store, retrieved the top 200 entries, disabled reranking, and followed mem0’s evaluation prompt setup. It used NeatMem v0.5.8 with BM25 on, entity extraction off, and thinking off.
| Configuration | Mean dedup calls | Mean merge calls | Mean total calls | Mean LoCoMo score |
|---|---|---|---|---|
| Add-Only | 0 | 0 | 1,400 | 90.68% |
| PointWise + Rewrite, threshold 0.8 | 2,018 | 576 | 3,994 | 90.27% |
| PointWise + Rewrite, threshold 0.4 | 23,985 | 1,801 | 27,186 | 89.13% |
| ListWise + Replace, threshold 0.4 | 5,021 | 0 | 6,421 | 89.52% |
| ListWise + Rewrite, threshold 0.4 | 4,997 | 366 | 6,763 | 90.47% |
| ListWise + Rewrite, threshold 0.8 | 1,263 | 183 | 2,846 | 90.75% |
The highest score exceeded Add-Only by just 0.07 percentage points. Kan describes the observed score band, 89.13% to 90.75%, as differences on the order of run-to-run variance. Meanwhile, PointWise + Rewrite at 0.4 averaged 27,186 total calls, compared with 6,763 for ListWise + Rewrite at 0.4. Those call counts are specific to this setup, not universal costs for implementing either strategy.
Free tools Windows power users keep installed
One-click scans. No signup required.
The threshold was a uniform NeatMem cosine-similarity threshold for this test. Thresholds used by different vendor implementations are not directly comparable, so these values should not be treated as product settings or transferable defaults.
Why a high score can hide a poor memory experience
LoCoMo accuracy measures whether the system can answer benchmark questions; it does not by itself establish that the underlying store is clean, current, or easy to maintain. With top-200 retrieval, duplicate or stale entries can coexist with a correct memory and still leave the answer model enough information to respond correctly. That can make a benchmark score look healthy even when updates are failing.
A similarity threshold can block a necessary update
In Kan’s Seattle-to-Austin example, the old and new location facts had reported similarity of roughly 0.71–0.75. At a candidate threshold of 0.8, PointWise did not send the contradictory pair to the language-model judge. Both the old Seattle fact and the newer Austin fact remained in memory. The failure occurred before the judge could make a decision: pairwise reasoning cannot resolve a conflict if retrieval excludes the relevant candidate.
Replacement can erase still-useful context
In a breakfast-routine example, wholesale replacement updated a coffee routine but lost the old detail that the person listened to Spanish podcasts. The new entry captured the changed routine without preserving that still-relevant preference. Rewrite instead merged the updated routine with the earlier information.
One-target updates can leave other stale entries behind
Kan’s tea example involved a new fact that invalidated multiple old entries. Single-target ListWise + Rewrite updated one item but left an old coffee-shop memory phrased as if it were still current. Multi-target ListWise + Rewrite judged the candidates together and used one merge call to combine the update with both stale items.
Why Kan recommends multi-target ListWise + Rewrite
For the protocol he tested, Kan’s preferred configuration is multi-target ListWise + Rewrite. Batch judgment can keep the number of deduplication calls lower than checking candidates one by one; multi-target handling can address several outdated entries at once; and Rewrite can preserve useful old detail that Replace would discard. Its 90.75% score was not meaningfully separated from other high-performing configurations given the reported run-to-run variation, so the recommendation rests on update behavior and call volume as well as benchmark accuracy—not on a clear accuracy win.
This is a recommendation for Kan’s reported setup, not proof that one strategy is best across products, models, or workloads. His experiment also does not establish that thresholds or call counts transfer to other memory systems.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Memory strategy depends on the task, not only on deduplication
EvoMemBench evaluates memory across two dimensions: whether information is kept within an episode or across episodes, and whether the stored content is knowledge or execution experience. Its authors report that retrieval-based approaches can be strong for knowledge-oriented demands, while procedural or long-term memory can help more on execution-oriented tasks when the stored experience matches the target decision process.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsThe same work identifies ways memory can hurt: irrelevant evidence may distract, compression may remove details needed to execute a procedure, and a procedure from a mismatched situation may transfer badly. The practical implication is to choose memory content and retrieval behavior for the task being supported, rather than treating deduplication or a single benchmark score as a complete measure of memory quality.
Read benchmark scores within their scope
LoCoMo, LongMemEval, and MemoryArena do not measure an interchangeable universal property called “memory quality.” The Memory Papers benchmark overview describes LoCoMo as covering up to 32 conversation sessions, LongMemEval as multi-session question answering with 500 questions, and MemoryArena as testing interdependent agent tasks. A result depends on task design, model, answer and evaluation setup, memory horizon, and retrieval configuration.
For context—not as a direct ranking against Kan’s six configurations—Redis reported 86.5% task-averaged accuracy for its Remis + Instruct strategy on LongMemEval and 42.4% on its Vertex Memory Bank test. These are vendor-reported results in Redis’s September 21, 2026 report. The report describes a 500-question benchmark, about 24 prior sessions per question on average, and a GPT-4o binary judge. Its systems and evaluation scope differ from Kan’s NeatMem LoCoMo experiment, so the percentages should not be compared as if they were scores on one common test.
Quick Recap
Practical takeaway for designing an agent memory
- Check whether candidate retrieval can surface relevant old entries before tuning the judgment prompt. A missed candidate cannot be reconciled downstream.
- Decide whether replacement is safe for the type of memory being stored. If old entries may contain still-valid details not repeated in a new fact, merging is less likely to discard them.
- Consider whether a new fact can invalidate several stored records. If so, a single-target update risks leaving contradictions behind.
- Evaluate more than answer accuracy: inspect stale and duplicate records, retained detail, update behavior, and model-call volume under the workload you actually care about.
- Treat benchmark thresholds and scores as setup-specific evidence. A result from one task, model, retrieval configuration, and memory horizon does not establish a universal best strategy.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →
