Crashes, 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 minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Published results do not establish one universally fastest or cheapest AI agent memory architecture: they use different workloads, configurations, and measurement boundaries. A 2026 repository run offers one useful, configuration-specific comparison, but it is not a universal ranking. The results below are attributed to their publishers rather than presented as a shared four-way benchmark.
What kinds of agent memory are being compared?
An architecture describes how a system represents, updates, and retrieves information; it is not the same thing as a product or framework. These four patterns help organize comparisons, but they should not be mistaken for a claim that every cited benchmark tested exactly one implementation of each pattern.
As an Amazon Associate I earn from qualifying purchases.
Vector or extraction memory
This pattern selects facts or other useful content for storage, then retrieves relevant items through similarity search or related search methods. AgentMemBench classifies Mem0 and LangMem as vector-based systems with strong LLM coupling. That coupling matters: extraction and retrieval behavior can depend on the model and prompts as well as the underlying index. AgentMemBench’s architecture overview describes its categories.
Temporal graph memory
A temporal knowledge graph represents entities and their relationships while preserving how those relationships change. The Zep authors describe Graphiti as a temporal graph engine that combines conversational and structured data and retains historical relationships. This representation is intended to support time-aware questions, but the architecture alone does not guarantee correct answers to them. The Zep paper explains the design.
#1 Best Overall
- EVOLUTION AMD RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Hierarchical, agent-managed memory
MemGPT’s virtual-context approach moves information among memory tiers, with the agent managing what remains in its active context. This puts agent decisions about what to retain, retrieve, or move at the center of the design. The MemGPT paper describes the approach.
File-backed or long-context retrieval
In a file-backed design, the agent searches stored conversation files using file operations. Letta describes a LoCoMo setup that combines semantic search and text matching. Its behavior therefore reflects the agent’s ability to use those tools, not only the contents of the files. Letta’s benchmark post documents that setup.
What published numbers can—and cannot—tell you
The clearest recent multi-system figures in the cited material come from the agent-memory-bench repository’s September 23, 2026 run. The repository reports a 419-turn workload using its stated harness and configurations. These are results for that run, not stable performance guarantees or a controlled comparison across every possible deployment.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
- Built for Local AI Development: AMD Ryzen AI Halo is designed for local AI development and inference, featuring 128GB unified memory and support for up to 200B parameter models to build and run intensive AI workloads locally.
- 128GB Unified Memory: Features 128GB LPDDR5x unified memory at 8000 MT/s with 256 GB/s memory bandwidth, providing a shared memory pool across the CPU, GPU, and NPU to support larger AI models.
- AMD Ryzen AI Max+ 395 Processor: Features 16 cores, 32 threads, and Zen 5 architecture, paired with AMD Radeon 8060S integrated graphics featuring 40 RDNA 3.5 compute units and an AMD XDNA 2 NPU with up to 50 TOPS.
- Linux AI Developer Platform: Purpose-built for Linux-based AI development with full AMD ROCm software support and preloaded tools, models, and workflows optimized for local AI development.
- Compact, Connected Design: Includes a 2TB M.2 SSD, 10GbE LAN, Wi-Fi 7, Bluetooth 5.4, USB-C connectivity, and HDMI 2.1b.
| Configuration named by the repository | Recall | Search p50 | Memory tokens | Ingest per turn |
|---|---|---|---|---|
| GoodMem vendor configuration | 57.6% | 754 ms | 504 | 0.28 s |
| Letta 0.11.7 | 52.2% | 318 ms | 503 | 0.37 s |
| Mem0 2.1.0 | 50.0% | 38 ms | 353 | 1.52 s |
| LangMem | 45.7% | 68 ms | 884 | 4.15 s |
| Zep/Graphiti | 37.0% | 163 ms | 212 | 3.65 s |
The repository’s run and configuration details are essential context for every number in this table. In this workload, Mem0 had the lowest reported search p50 and memory-token count, while GoodMem had the highest recall and the highest search p50. Mem0 also had a higher reported ingest time per turn than GoodMem, Letta, and Zep/Graphiti. Those trade-offs make it misleading to call one row simply “best.” Memory tokens are not a complete token bill, and search p50 does not include every part of an agent’s response cycle.
Other published figures use different evaluations and should be read separately:
- LoCoMo: Letta reported 74.0% accuracy for a file-backed agent using GPT-4o mini with constrained tool rules. In the same post, Letta compared that with a reported 68.5% score for Mem0’s graph variant while discussing evaluation challenges. These are vendor-published figures, not an independent, shared head-to-head test. Letta’s post gives the comparison context.
- DMR: The Zep paper authors reported 94.8% for Zep versus 93.4% for MemGPT on this benchmark. This is a result for the paper’s stated setup, not a general ordering of graph and hierarchical memory. The paper describes the comparison.
- LongMemEval: The Zep paper authors reported up to an 18.5% accuracy improvement and 90% lower response latency relative to the baseline implementations in their comparison. “Up to” and the specified baseline context matter; these figures do not establish the same gain against every memory design or configuration. The paper provides the comparison.
Why latency and token rankings are easy to misread
“Latency” can mean a database search, a complete memory-tool cycle, ingestion, or end-to-end answer time. Those measures answer different questions. A system with a fast search call can still take longer overall if it spends more time extracting, making tool calls, or generating an answer. Report ingestion separately from retrieval and answer generation.
Rank #3
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Likewise, “tokens” may mean only retrieved memory or the entire model request, including instructions, conversation history, and the answer-generation prompt. A low retrieved-memory count can reduce one part of a request without proving lower total usage or monetary cost. Any comparison should state exactly which tokens it counts.
AgentMemBench identifies six useful evaluation axes: write efficiency, retrieval quality, scalability, temporal consistency, isolation and privacy, and LLM portability. Its documented metrics include write and read latency, recall and omission rate, latency and exact-canary recall at different fact counts, stale-fact and update behavior, cross-user leakage, deletion completeness, and backend portability. Its benchmark documentation lays out the axes. A separate Agent Memory Benchmark repository documents common harnesses and comparisons involving systems such as Mem0, Letta, Graphiti, and LangMem, with datasets including LoCoMo and LongMemEval.
How to run a fair four-way comparison
- Pin what is being tested. Record each system’s exact version or commit and whether it is hosted or self-hosted. Include the storage and search configuration, model and model version, embedding model, decoding settings, prompts, tool rules, and judge model. Without these details, another team cannot tell whether a result follows from the architecture or from its setup.
- Use the same workload and answerability rules. State the dataset, conversation length, number and types of questions, and whether the required facts are actually present in the conversation. Include both straightforward lookups and questions requiring updates, time reasoning, or multiple facts.
- Separate the timing stages. Measure ingestion, memory search, the complete retrieval or tool cycle, and end-to-end answer time independently. Report p50 and p95 for latency, with a clear start and stop boundary for each measure.
- Define the token boundary. Report retrieved-memory tokens and total model-request tokens separately. If reporting monetary cost, state which model or service prices and which calls are included rather than treating token counts as prices.
- Score more than exact answers. Report recall or accuracy alongside omissions, incorrect answers, and abstentions. Repeat runs where system behavior can vary and include uncertainty or run-to-run variation; do not treat a single point estimate as conclusive.
- Test operations and isolation. Ask update and temporal questions, check whether obsolete facts persist, probe unanswerable cases, test cross-user leakage, and verify deletion. These are distinct from ordinary recall and can expose important failures that an aggregate score hides.
- Publish enough to reproduce the result. Release the workload or its specification, configurations, scoring rules, and artifacts where permitted. Pin repository release tags or commits: a benchmark name alone does not identify the code or settings used for a particular result.
Which failure modes should a benchmark expose?
The following are failure cases to test, not findings established for any one architecture by the cited figures. Recording where the failure occurs helps avoid blaming storage for a problem caused by the agent, model, or scoring procedure.
| Test case | What to record |
|---|---|
| The relevant fact is never stored. | Check the source conversation against stored records; investigate extraction or ingestion before attributing the miss to retrieval. |
| The fact is stored, but the agent does not search for it or invoke the memory tool. | Record tool calls and prompts so tool-selection failure is distinguishable from an index miss. |
| Search returns a similar but incorrect fact, and the agent answers confidently. | Inspect retrieved records, the generated answer, and whether the agent could abstain; distinguish retrieval error from answer-generation error. |
| A newer value fails to replace or qualify an obsolete one. | Use update and time-sensitive questions, then inspect stored versions and the answer’s temporal grounding. |
| A temporal or multi-hop question fails although its component facts are present. | Check whether the missing step is retrieval, relation or time reasoning, or answer synthesis. |
| Retrieval brings back excess context. | Measure irrelevant retrieved content, memory tokens, latency, and whether additional context changes answer quality. |
| One user’s records appear in another user’s results, or deletion leaves records behind. | Test tenant isolation and deletion completeness directly rather than inferring privacy from recall scores. |
| Ingestion is slow or costly even when search is fast. | Report ingestion time and its model or processing calls separately from search latency. |
| The score changes because the benchmark or judge misclassifies an answer. | Inspect question answerability, scoring rules, and judge behavior; validate disputed cases rather than treating the aggregate as ground truth. |
How to interpret the evidence when choosing a design
Start with the workload you need to support, then compare configurations using the same model, questions, timing boundaries, and scoring rules. If search latency is the priority, the 419-turn repository run gives a result for its Mem0 2.1.0 configuration—not proof that vector memory is always fastest. If time-aware relationships matter, a temporal graph is designed to represent them, but it still needs direct testing on your update and temporal questions. If context management or searchable conversation files fit your agent, hierarchical and file-backed approaches merit evaluation under the same tool policy and model.
Letta’s interpretation is that “The quality of an agent’s memory often depends more on the underlying agentic system’s ability to manage context and call tools than on the memory tools themselves.” That is the company’s interpretation, not a neutral consensus, but it points to a crucial benchmark-design issue: an agent that fails to store, search, or correctly use a fact can make a memory backend appear worse than it is. Letta’s discussion provides the quotation’s context.
Quick Recap
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.
