Local-first context + evidence
MemStrata is a local-first product suite for coding context, conversational memory, deterministic operational answers, and AI-agent governance.
Keep sensitive working data close, route only what is needed, and inspect the evidence behind every result.
A bounded pulse moves through the strata unless reduced motion is requested. The labels remain readable without animation.
Products
Four products share an evidence spine. They are not four skins over Coder.
Coder
Route focused project context into coding tools and measure each visible user turn.
For: Developers and teams using AI coding tools who need current project context without sending a whole repository on every prompt.
linux-x64 · PyPI 0.6.32
Conversational
Recall conversations and documents with temporal context, citations, OCR, and optional visual retrieval.
For: Knowledge and support teams that need private memory over chats and documents, with visible evidence.
linux-x64 · PyPI 0.7.5
Structured
Ask operational data questions and receive deterministic, source-backed answers, or an honest refusal.
For: Operations and data teams that need current-state, count, owner, history, and as-of answers with witnesses.
linux-x64 · PyPI 0.7.1
Governance
Enroll AI systems and agents, apply policy gates, manage approvals and incidents, and retain verifiable evidence.
For: AI risk and governance teams that need inventory, allow/review/block decisions, and an evidence chain.
linux-x64 · PyPI 0.7.1
Coder connects through MCP, an OpenAI-compatible harness, or the VS Code extension. The extension also loads in VS Code shims such as Cursor and Windsurf. Native tab-complete and some closed Composer paths are partial, and we label them that way.
Install once, then connect the editor you already use.
The VS Code extension runs in VS Code and in forks that load VS Code extensions, including Cursor and Windsurf.
Point MCP or an OpenAI-compatible harness at the local Coder runtime.
Sources → Local index/state → Retrieval or deterministic decision → Reader/action → Evidence receipt
Local-first project context, routing, MCP and tool setup, and measured AI-coding telemetry.
Evidence-backed conversational memory across chats, documents, and explicitly connected sources.
Deterministic answers, history, and proof from governed operational records.
AI-system and agent inventory, policy decisions, approvals, incidents, alerts, and tamper-evident evidence.
Each item links to method or product evidence. Coder savings are not the identity of the suite.
81.4% input-token avoidance (n=4 controlled user turns)
81.4% input-token avoidance observed in a four-turn controlled Coder E2E test against a one-time full-repository session-start counterfactual. Results vary by repository, client, model, and task.
Methodology and sample size
Cite sealed published evaluation results only when runner, sample, reader, scorer, and artifact hash are public. Incomplete competitor arms are not compared here.
A supported answer returns a source witness. An unsupported or predictive question is refused.
Allow, review, or block with an evidence chain. This is not a legal compliance certification.
| Need | Product | Output | Model required? |
|---|---|---|---|
| Focused coding context and savings telemetry | Coder | Context plus per-turn measurements | User-selected coding model |
| Memory across chats and documents | Conversational | Evidence-backed model answer | Local or BYOK reader |
| Verified operational-record questions | Structured | Deterministic answer or refusal plus proof | No model for core answers |
| Agent policy and audit evidence | Governance | Policy decision plus evidence | No hidden-reasoning invention |
Privacy-critical content and indexes remain on the device by default. OAuth, signed entitlements, device registration, and bounded private decisions use MemStrata services. A user-selected cloud model or connector receives only the data required for that explicit operation. Credentials are stored in the operating-system credential vault, not in page content or analytics. Customer content is not used for training unless the customer separately opts in.
Published papers are research measurements, not product SLAs. Draft papers stay labelled as draft.
Paper 1 · Published
Deterministic supersession that RAG cannot match by construction
arXiv:2606.26511Paper 2 · Published
130 marker-free atomic transitions from 707 real SWE-bench GitHub fixes
arXiv:2608.20685The 180-day trial begins after successful product activation. Product functionality remains enabled during the trial. MemStrata-managed email is limited to one message per signed-in user per UTC day. Reinstalling does not restart access. A paid entitlement is required after the trial.
Keep project context local, route only what a coding tool needs, and measure each visible user turn.
Search chats and documents with temporal context and visible evidence, including OCR when scans are in play.
Ask current-state and as-of questions of governed records and receive a witness, or an honest refusal.
Inventory agents, apply allow/review/block gates, and keep a hash-linked evidence chain.