A changed decision
Illustrative exampleWhen is the project review?
Friday.
The Tuesday note is still in the record. A later note moves the review to Friday.
New Conversational 0.8.0 is out
AI answers you can check.
MemStrata keeps your chats, documents and project history in an index on your own computer. Ask a question and see the exact passages behind the answer.
Free for 180 days after activation. Windows, Linux and Apple silicon.
Measured in the open
95%LongMemEval-S source-aware475/500 · Qwen 3.8 27B Q4Author evaluationDenominator.LongMemEval-S, 500 questions, CL1 24k, same retained answers from Qwen 3.8 27B Q4, full-source GPT 5.5 regrade.When.Author evaluation, source-aware regrade recorded 2026-09-22Limitations.Source-aware regrade of answers already produced. The reader was not run again.The reference-aware score on the same answers is 463 of 500 (92.6%).Development-exposed research configuration, not a shipping product measurement.Not a product SLA.MethodsReproduction
90.91%LoCoMo source-aware1400/1540 · Qwen 3.8 27B Q4Author evaluationDenominator.LoCoMo 1540 scored QA (categories 1-4), CL1 24k, same retained answers from Qwen 3.8 27B Q4, GPT 5.5 source-aware regrade.When.Author evaluation, source-aware regrade recorded 2026-09-23Limitations.Source-aware regrade of answers already produced. The reader was not run again.The reference-aware score on the same answers is 1205 of 1540 (78.25%).Development-exposed research configuration, not a shipping product measurement.Not a product SLA.MethodsReproduction
85.4%LongMemEval-M reference-aware427/500 · Qwen 3.8 27B Q4Author evaluationDenominator.LongMemEval-M, 500 questions, CL1, reader Qwen 3.8 27B Q4, GPT 5.5 reference-only judge. 179/200 plus 248/300.When.Author evaluation, LME-M full 500 graded 2026-09-27Limitations.Reference-aware only. The histories are too long for the judge to read as source, so no source-aware score was produced.Reader output failures are scored 0 and stay in the 427 of 500.The remainder (248/300) scored lower than the first sample (179/200).Development-exposed research configuration, not a shipping product measurement.Not a product SLA.MethodsReproduction
Author-run, development-exposed research using one automated judge model, GPT 5.5—not independent replication or a held-out customer evaluation. The source-aware headlines regrade the same answers with sources visible to the judge. Their reference-aware scores are 92.6% (463/500) on LongMemEval-S and 78.25% (1205/1540) on LoCoMo. These are not new reader runs or product guarantees. Compare the grading methods.
The research reader was Qwen 3.8 27B Q4. The built-in chat suggests that model but does not require it: you can choose a local Ollama model, a local OpenAI-compatible server, or a cloud provider you configure, and the published accuracy figures apply only to the measured Qwen configuration. Weights are not bundled. Claude Code and the VS Code extension let the AI you already use read the evidence instead.
How it works
Three steps from your files to an answer you can check.
- STEP 01
Connect sources
Prepare private memory, then add chats, notes or documents.
- STEP 02
Choose the answer model
Setup lets you choose the answer model used for Standard and Deep.
- STEP 03
Ask and open the evidence
Ask in the chat. Standard and Deep use the same index. Open Why this answer and read the passages behind the answer.
What happens on your computer, and what is sent to services
Indexes and source processing run on your device. Account, licensing, and optional connectors use MemStrata or destination services. The built-in Conversational chat answers with the model you choose: with a local model the question and evidence stay on your computer, and with a cloud provider they go to that provider after you consent. If you send evidence to a cloud model or an MCP client, the question and selected passages leave the device for that provider. Credentials stay in the operating-system credential vault. Customer content is not used for training unless the customer separately opts in. Conversational 0.8.0 also makes two requests to public services whether or not you use a cloud provider. At startup and every 24 hours it downloads OpenRouter’s public model price list to estimate cloud-provider costs, and when you open the dashboard it requests the US-dollar to Canadian-dollar exchange rate from the Bank of Canada’s public data service to show subscription amounts. The requests contain no API key, question, source content or MemStrata account identifier, but they reveal your IP address to OpenRouter and the Bank of Canada. There is no setting to turn them off in 0.8.0. We plan to make both happen only when a cloud provider is configured in a later release.
See the idea
When a fact changes, the older record stays openable.
A deadline moves. A count spans many conversations. A detail was never mentioned. Step through three short examples and open the sources behind each answer.
Four products
One memory layer. Four jobs.
Conversational, Coder, Structured and Governance each do one job well. Start with the one that matches your work.
Start here
Conversational
Ask about imported documents and chats with inspectable evidence
- Text and PDF sources
- Inspectable evidence
For developers
Coder
Supply relevant repository context to a supported coding tool
- Selected project files
- Context receipt
For operations
Structured
Answer supported current, history and change questions over records
- Record identity and time
- What changed
For AI teams
Governance
Observe connected agent actions and use supported controls
- Connected-host event
- Policy decision
Measured in the open
Read our published research on arXiv. Every benchmark result on this site keeps its denominator, judge and method label. Research results, measured under the stated conditions.
- arXiv 2606.26511Temporal Validity in Retrieval Memory: Eliminating Stale-Fact Errors for AI Agents over Evolving Knowledge
- arXiv 2608.20685Temporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory over GitHub Fixes
- arXiv 2610.07715When Old Facts Return: Re-Reads, Reverts, and the Limits of Temporal Memory
- arXiv 2610.05343MemStrata: 95% and 90.91% Source-Aware Accuracy on LongMemEval-500 and LoCoMo-1540 with a Local Qwen 3.8 27B Q4_K_M Reader
Clear about data
Indexes and source processing run on your device. Account, licensing, and optional connectors use MemStrata or destination services. The built-in Conversational chat answers with the model you choose: with a local model the question and evidence stay on your computer, and with a cloud provider they go to that provider after you consent. If you send evidence to a cloud model or an MCP client, the question and selected passages leave the device for that provider. Credentials stay in the operating-system credential vault. Customer content is not used for training unless the customer separately opts in. Conversational 0.8.0 also makes two requests to public services whether or not you use a cloud provider. At startup and every 24 hours it downloads OpenRouter’s public model price list to estimate cloud-provider costs, and when you open the dashboard it requests the US-dollar to Canadian-dollar exchange rate from the Bank of Canada’s public data service to show subscription amounts. The requests contain no API key, question, source content or MemStrata account identifier, but they reveal your IP address to OpenRouter and the Bank of Canada. There is no setting to turn them off in 0.8.0. We plan to make both happen only when a cloud provider is configured in a later release.
Read the trust pageStart with one source and one question.
Install Conversational, connect a folder of notes or an exported chat, and open the evidence behind your first answer.