01
What is temporal validity in AI agent memory?
It is knowing which facts are still true after knowledge evolves. MemStrata retires contradicted values with a deterministic supersession rule — no similarity threshold, no LLM on the read path.
开源 · 本地优先
别再把整个代码库 每次都发给 AI。
MemStrata 是面向 AI 编程工具的本地上下文层。中位用户：输入 token 比 RAG 少 42%，每月省回 $14。
免费开源内核。若 Pro 的节省不超过订阅费，我们退款。
仪表盘 — 金钱标签页
live
Saved this month ▲ +12% vs last
$14.62
Tokens vs. RAG 42%
Recall 87%
Precision 64%
24h survival 73%
Works with the tools you already use
Cursor Claude Code VS Code Windsurf Cline JetBrains Aider Copilot Cursor Claude Code VS Code Windsurf Cline JetBrains Aider Copilot
Capabilities
Built like a research product. Feels like magic.
Bento layout of the moat — hover any tile for a spotlight interaction.
Temporal validity ~0% stale-fact error
RAG has no model of time. We do.
Deterministic (subject, relation, object) supersession retires stale facts in a bi-temporal ledger — AUROC 0.59 proves cosine cannot tell contradiction from duplicate.
Explore Local-first 100% local index
Your code stays here
SQLite + DuckDB on-device. License check is a signed JWT only.
Explore Latency ~2.1s retrieval
Embedding floor, not LLM tax
No model on the read path. ~2.1s vs 16–18s for rerankers.
Explore Token economics 42% vs naive RAG
Median 42% fewer input tokens
Compress context per turn. Watch dollars tick up on the local Money tab — money-back if Pro savings fall short.
Explore Integrations 11+ tool surfaces
Cursor · Claude Code · VS Code · more
MCP, harness, and extension paths — honest coverage matrix, no oversell.
Explore Research P1–P5 portfolio
Five papers. One temporal spine.
From arXiv temporal validity through SWE-bench longitudinal, value-change ceilings, detect-and-flag logic bulk, and world-knowledge generalization.
Explore Scrollytelling
Scroll the story of temporal memory
Sticky visual layer · glass panels · research-backed answers
02
Why does RAG serve stale facts?
When a fact changes, old and new embeddings sit next to each other (cosine AUROC ~0.59 for contradiction vs duplicate). RAG retrieves both and has no structural way to choose the current value.
03
How is MemStrata different from Mem0 or Graphiti?
Same-stack local 7B tests show MemStrata’s CERTAIN spine at 1.000 on supersession and TEMPO axes. We optimize for never confidently wrong on evolved knowledge — not just long-dialogue recall leaderboards.
04
Does this cut AI coding token costs?
Yes. Median users see ~42% fewer input tokens vs naive RAG, with live savings on a local dashboard. Pro includes a money-back guarantee if measured savings fall short.
Workflow
工作原理
三步。不到 60 秒即可看到首次节省。
在本地安装 MemStrata
pipx install memstrata-pro memstrata init
约 60 秒，完全本地
照常编程
MemStrata 构建你代码库的图谱，并在每次 AI 请求时自动在后台压缩上下文。
看着你的节省增长
实时仪表盘显示每轮节省的 token 与美元，并附四项指标的明细。
Telemetry
我们度量为你节省了多少
四个数字都在你的本地仪表盘上。没有“最高 X%”——只有你的真实遥测。
42%
相比 RAG 节省的 token
活跃用户中位数。你的仪表盘显示你的精确数字。
Computed per AI turn: actual tokens sent ÷ what a naive top-5 BM25+embedding RAG pipeline would have sent. 42% is the cross-user median.
Full methodology →
87%
检索召回率
每轮中位数——AI 引用且存在于我们上下文中的实体。
After each turn, MemStrata parses the AI response for referenced symbols/files and checks how many were in the context it provided.
Full methodology →
64%
检索精确率
每轮中位数——我们提供且 AI 实际使用的实体。
High precision means focused context, not padding. Low precision shows up as higher token costs — caught and surfaced in the dashboard.
Full methodology →
73%
24 小时编辑留存
每会话中位数——24 小时后仍在仓库中的 AI 编辑。
Tracks whether AI-suggested edits survive the next day of development. A proxy for suggestion quality — bad suggestions get reverted quickly.
Full methodology →
Integrations
能与你的技术栈协作吗？
诚实的兼容矩阵。附带说明——我们不夸大封闭生态。
| Tool | MCP | Harness | Extension |
|---|---|---|---|
| Cursor | — | ||
| Windsurf | — | ||
| VS Code | — | ||
| Claude Code | — | ||
| Cline | — | ||
| Continue.dev | |||
| JetBrains AI | — | — | |
| GitHub Copilot Chat | — | ||
| Aider | — | — | |
| Codex CLI | — | — | |
| Zed | — |
No tools match your search.
Privacy
你的代码永不离开你的机器
许可服务器仅通过签名令牌验证你的订阅——别无其他。
Your machine — all local
AI tool
Claude Code, Cursor, VS Code…
Core
MemStrata harness
localhost:8080
compresses context
Your LLM provider
Anthropic, OpenAI, Ollama…
Local SQLite + DuckDB
· index.db · telemetry.db — your code stays here
License server
memstrata.dev/lic
Pricing
简单、诚实的价格
永久免费且开源。付费档位增加主动式 harness 与退款保证。
免费
开源 $0 永久
MIT 许可的开源内核。永远属于你。
- Open-source MIT core
- Local MCP server
- Local dashboard
- Browser extension — chat capture
- Community support (GitHub)
最受欢迎
Pro
$14.99 CAD /mo
＋税
主动式上下文 harness ＋ V6 记忆引擎。
- Everything in Free
- Active context harness — all AI tools
- VS Code + browser augmentation
- V6 memory engine — verified retrieval
- Money tab — full savings metrics
- 60-day money-back guarantee
Pro+
$16.99 CAD /mo
＋税
包含 Pro 全部，外加 V7 工具自动生成。
- Everything in Pro
- V7 tool autogeny — auto tool gen + JIT routing
- Earliest access to new engines
- Priority support
永久使用免费开源内核——无需账户。 完整对比 →
Research
Backed by science — not vibes
Temporal validity, marker-free evaluation, and a five-paper program on agent memory under knowledge evolution.
0.59
Cosine AUROC
~0%
Stale-fact error
0.95–1.0
Evolving accuracy
~2.1s
Retrieval latency
1 Published arXiv preprint · cs.CL / cs.AI
Temporal Validity in Retrieval Memory: Eliminating Stale-Fact Errors for AI Agents over Evolving Knowledge
Deterministic supersession that RAG cannot match by construction
RAG gives agents access to accumulated knowledge but has no model of time. When a fact changes, cosine similarity surfaces both stale and current values nearly equally (AUROC 0.59 for contradiction vs duplicate). MemStrata stores facts like RAG, then retires contradicted values with a deterministic (subject, relation, object) supersession rule in a bi-temporal ledger — no similarity threshold, no LLM on the read path. Across six local benchmarks with a 7B model, MemStrata ties RAG on static knowledge and reaches 0.95–1.00 accuracy on evolving knowledge where RAG reaches 0.20–0.47. Stale-fact-error drops from 15–40% (RAG, when forced to answer) to ~0%.
- Cosine AUROC 0.59 — similarity cannot separate contradictions from duplicates
- Evolving knowledge accuracy 0.95–1.00 vs RAG 0.20–0.47
- Stale-fact-error ~0% vs RAG 15–40% when required to answer
- Retrieval ~2.1s (embedding floor) vs ~16–18s LLM-rerank baselines
temporal validitystale-fact errorbi-temporal ledgerRAG
Blog
All 10 posts → Explainers that drive SEO & understanding
Product + research
8 min read Paper 1, 2
Cut AI Coding Token Costs with Local Context Compression (Without Sending Your Repo Upstream)
MemStrata is the local context layer for AI coding tools: median 42% fewer input tokens vs naive RAG, money-back on Pro, and research-backed temporal memory so cheaper context is still correct context.
token costs AI coding local-first
Competitive analysis
11 min read Paper 1, 4, 5
MemStrata vs Mem0, Graphiti, and Letta: Same-Stack Comparison on Temporal Axes
Apples-to-apples local 7B evaluation: MemStrata’s deterministic supersession hits 1.000 on MemArch supersession/poisoning/TEMPO axes where flat RAG and several agent-memory systems still serve stale values.
Mem0 Graphiti Letta
Paper 2
8 min read Paper 2, 1
SWE-bench Longitudinal: Temporal Memory Beats RAG on Real Code Histories
On 130 marker-free scenarios extracted from SWE-bench Lite/Verified buggy→fixed histories, MemStrata temporal_v6 hits 0.908 accuracy vs ~0.57 RAG — with stale-fact-error 0.023 vs 0.262.
SWE-bench code agents benchmarks
Ship more. Spend less on tokens.
Install in under a minute. Keep coding the way you already do. Watch the savings compound — with research-backed temporal memory underneath.