Benchmark evidence
Context,
under pressure.
Two public evaluations test long-term recall and repository-level software engineering with measurable outcomes.
Two benchmarks. Two different questions.
Memory quality matters twice: first when an agent must recover what happened, then when it must use context to finish real work.
Can the system remember a long-running conversation?
It measures direct recall, temporal reasoning, multi-step reasoning, and context grounded open-domain answers.
Explore the memory testCan an agent turn context into a working patch?
It gives Codex a real repository and issue, then grades the resulting patch with executable tests.
Explore the coding testLoCoMo tests long-term recall.
The public dataset contains long, multi-session conversations. PowerContext is evaluated on the answerable question set, where facts can be separated by many sessions.
- Conversations
- 10
- Scored questions
- 1,540
- Question types
- 4
Single-hop
Recover one fact from the conversation history.
Temporal
Reason about dates, order, and duration across sessions.
Multi-hop
Connect several facts before producing an answer.
Open-domain
Combine conversation evidence with general knowledge.
One run, three ways to carry context.
Switch metrics to compare PowerContext, PowerMem, and placing the entire conversation in the prompt.
Higher is better.
Lower is better.
Lower is better.
SWE-bench Pro tests whether context changes the patch.
Each task starts from a real codebase and issue. Codex edits the repository, and the official task tests decide whether the patch resolves the problem.
The public score index.
Put the published field next to PowerContext, while keeping the test rig visible.
Data checked August 31, 2026LoCoMo scores on all 1,540 questions
Only results that explicitly disclose the complete 1,540-question scope are included. Reader, judge, and answer-policy differences remain visible.
| Rank | System | Score | Published protocol | Evidence |
|---|---|---|---|---|
| 1 | Zep | 94.70% | 1,540 questions; GPT-5.4 reader and judge | Vendor run |
| 2 | EverMemOS | 94.50% | 1,540 questions; lenient shared harness; precomputed retrieval | Third-party run |
| 3 | XMDB | 93.20% | 1,540 questions; internally verified | Vendor run |
| 4 | TrueMemory Pro | 93.00% | 1,540 questions; three-run mean; lenient judge | Open harness |
| 5 | Mem0 | 92.50% | 1,540 questions; latest vendor evaluation rig | Vendor run |
| 6 | PowerContext | 90.78% | 1,540 questions; 1,398 correct; topical judge | Project run |
| 7 | Honcho | 89.90% | 1,540 questions; current full-system result | Vendor run |
| 8 | Dakera | 88.20% | 1,540 questions; single pass; no LLM reranker | Reproducible vendor run |
| 9 | PowerMem | 87.79% | 1,540 questions; historical project run | Project run |
| 10 | Memvid | 85.65% | 1,540 questions; GPT-4o reader; lenient judge | Open harness |
| 11 | Genesys | 85.55% | 1,540 questions; ten-run mean; frozen Mem0 protocol | Certified vendor run |
| 12 | Engram | 84.50% | 1,540 questions; shared lenient harness | Third-party run |
| 13 | MemHQ | 83.20% | 1,540 questions; gpt-4o-mini; partial answers accepted | Open harness |
| 14 | Logica Mind | 72.50% | 1,540 questions; Mem0 paper protocol | Open harness |
| 15 | Supermemory | 65.40% | 1,540 questions; shared lenient harness | Third-party run |
Every official public entry
Scale currently publishes 25 model runs, not 30 memory products. Official Rank (UB), resolve rate, confidence interval, and harness marker are preserved below.
| Official rank | Model | Resolve rate | Provider | Harness |
|---|---|---|---|---|
| 1 | Muse Spark 1.1* | 61.50%±3.10 | Meta | mini-swe-agent |
| 1 | gpt-5.4 (xHigh)* | 59.10%±3.56 | OpenAI | mini-swe-agent |
| 3 | Muse Spark* | 55.00%±3.60 | Meta | mini-swe-agent |
| 3 | claude-opus-4-6 (thinking)* | 51.90%±3.61 | Anthropic | mini-swe-agent |
| 5 | gemini-3.1-pro (thinking)* | 46.10%±3.60 | mini-swe-agent | |
| 5 | claude-opus-4-5-20251101 | 45.89%±3.60 | Anthropic | Scale run |
| 5 | claude-4-5-Sonnet | 43.60%±3.60 | Anthropic | Scale run |
| 5 | gemini-3-pro-preview | 43.30%±3.60 | Scale run | |
| 5 | claude-4-Sonnet | 42.70%±3.59 | Anthropic | Scale run |
| 10 | gpt-5-2025-08-07 (High) | 41.78%±3.49 | OpenAI | Scale run |
| 10 | gpt-5.2-codex | 41.04%±3.57 | OpenAI | Scale run |
| 10 | claude-4-5-haiku | 39.45%±3.55 | Anthropic | Scale run |
| 10 | qwen3-coder-480b-a35b | 38.70%±3.55 | Alibaba | Scale run |
| 14 | minimax-2.1 | 36.81%±3.55 | MiniMax | Scale run |
| 14 | gemini-3-flash | 34.63%±3.55 | Scale run | |
| 16 | gpt-5.2 | 29.94%±2.15 | OpenAI | Scale run |
| 16 | kimi-k2-instruct | 27.67%±3.25 | Moonshot | Scale run |
| 18 | qwen3-235b-a22b | 21.41%±2.25 | Alibaba | Scale run |
| 19 | gpt-oss-120b | 16.20%±2.67 | OpenAI | Scale run |
| 19 | deepseek-v3p2 | 15.56%±2.63 | DeepSeek | Scale run |
| 21 | gemma-3-27b-it | 11.38%±2.15 | Scale run | |
| 21 | llama3-1-405b-instruct | 11.18%±2.15 | Meta | Scale run |
| 21 | glm-4.6 | 9.67%±2.15 | Z.ai | Scale run |
| 24 | llama4-maverick-17b-instruct | 5.24%±1.24 | Meta | Scale run |
| 25 | codestral-2405 | 1.51%±1.51 | Mistral | Scale run |
Read each result for the question it answers.
The two evaluations share a context theme, but their inputs, outputs, and graders are intentionally different.
| Evaluation dimension | LoCoMo | SWE-bench Pro |
|---|---|---|
| What is tested | Long-term conversational recall and reasoning | Repository-level issue resolution |
| Input | Multi-session dialogue history and a question | A repository, an issue, and a clean task environment |
| Output | A grounded natural-language answer | A code patch |
| Primary score | Judge-rated answer accuracy | Official executable tests passed |
Evidence and methodology
Follow the dataset, paper, harness, and published PowerContext figures from the original sources.
- PaperEvaluating Very Long-Term Conversational Memory of LLM Agents
The ACL 2024 paper that defines LoCoMo and its long-term memory tasks.
- Datasetsnap-research/locomo
The public ten-conversation dataset and annotations.
- Benchmarkscaleapi/SWE-bench_Pro-os
The public benchmark repository and official evaluation path.
- HarnessPowerContext evaluation console
The pinned dataset, OFF and ON arms, isolated runner, and reporting contracts.
- ResultsPublished PowerContext benchmark figures
The current project README values used on this page.
Inspect the system behind the scores.
PowerContext is open source. Review the implementation, evaluation harness, and contracts directly.