Measure Selected Coding Context Without Sending the Whole Repository Upstream
MemStrata Coder routes selected project context into coding tools and measures each visible user turn. The 81.4% input example is a four-turn controlled test, not a universal savings rate.
By Neeraj Yadav. MemStrata is built by Neeraj Yadav.
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 Coder attacks both problems as a local-first coding-context product. It measures input, cache, and output per visible user turn. Research on temporal validity is published separately and is not a production-cohort result.
Local-first by architecture
Indexes and telemetry stay on your machine. The harness listens on localhost. Account, licensing and device registration still use MemStrata hosted services. Selected context can go to the coding tool or model you connect. MemStrata does not use your repository to train a MemStrata model. See Trust.
There is no automatic end-of-cycle credit note on current marketing offers. The Coder Savings Guarantee applies only when it is shown and accepted at direct Stripe checkout. Public prices are not published yet. Full terms: Returns, Refunds, Trials and Cancellation.
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. The earlier scope and structural-staleness studies, now merged into Paper 3, keep the discussion explicit about logic bulk — flag structure that cannot be triple-extracted instead of inventing false certainty.
Read the science at arXiv:2606.26511, then install Coder and inspect the context receipt for one visible user turn. The 81.4% input-token example is a four-turn controlled test. See the methodology.
Previous: Comparing MemStrata, Mem0, Graphiti and Letta: what the evidence can tell you Next: How to Ask Questions Across Your Documents and Past Conversations