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.
Cheaper context is useless if it is wrong
AI coding tools bill you for tokens. The naive fix is aggressive retrieval truncation — which often drops the one file that mattered. The sophisticated failure mode is more subtle: you retrieve the right topics but the wrong version of a fact, then pay again to debug the wrong advice.
MemStrata attacks both problems. As a product, it is a local context layer for Cursor, Claude Code, VS Code, Windsurf, Cline, and more — median users see 42% fewer input tokens vs a naive top-k RAG baseline, with savings shown on a local dashboard. As research, it adds temporal validity so compressed context stays current.
Local-first by architecture
Indexes and telemetry stay on your machine (SQLite + DuckDB). The harness listens on localhost. The only outbound license check is a signed JWT — no code, prompts, or responses leave for “training.” That is why security-conscious teams can adopt compression without adopting a new data plane.
Pro tiers add the active harness, V6 memory engine, and a money-back guarantee: if measured savings fall short of subscription cost, the difference is credited automatically.
How research shows up in the product
The V6 memory engine ships the Paper 1 supersession spine: deterministic retire of contradicted facts, bi-temporal ledger, marker-free evaluation ethos. Paper 2’s SWE-bench longitudinal work is why we care about real repo histories, not just synthetic templates. Papers 3–4 keep the product honest about logic bulk — flag structure that cannot be triple-extracted instead of inventing false certainty.
Read the science at arXiv:2606.26511, then install locally and watch the Money tab.