17 posts on temporal validity, stale-fact errors, software histories, and local coding context, grounded in the paper series and arXiv:2606.26511.
Explore the research portfolio →Featured
Standard mode uses 8,192 evidence tokens; Deep uses 24,000. These are source budgets, not model context windows or the R454 7,000-token research budget. Screenshots were captured on 0.7.13.
Conversational indexes sources on the device. You connect your own AI. Account and licensing still call MemStrata hosted services. Screenshots are from the historical 0.7.13 walkthrough.
A citation is not a proof. Three small examples show an irrelevant source, an outdated source, and an unsupported conclusion. Inspect the passage, not only the link.
A historical Conversational walkthrough captured on 0.7.13: import demo notes and inspect the evidence. Use the install selector for the current release.
How Coder connects to VS Code, VS Code shims such as Cursor and Windsurf, JetBrains IDEs, and terminal agents including Claude Code, Codex CLI, Aider, and OpenCode via MCP, harness, or extension.
arXiv:2608.20685 validates deterministic supersession on 130 marker-free atomic transitions from 707 SWE-bench GitHub issues. MemStrata reaches 0.91 accuracy versus RAG 0.57-0.59, with stale-fact error near 0 versus 36-38%.
LongMemEval is not one number. A useful comparison states the dataset, reader, evidence budget, judge and failure handling. The public MemStrata R454 figure is 84.0% (420/500) judged by GPT 5.5.
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.
A practical guide to comparing AI memory systems, with MemStrata's scoped LongMemEval and LoCoMo results against Letta and a frozen Zep snapshot.
Oracle experiments on synthetic world-knowledge facts hit 1.000 accuracy / 0.000 stale in that protocol. Remaining error is extraction. This note is not the Paper 4 conversational 27B evaluation.
You cannot supersede logic you cannot extract. The earlier structural-staleness design marks scope-level code changes and warns at retrieval — without faking understanding or breaking the deterministic moat.
Across 707 SWE-bench fixes, deterministic temporal memory natively covers ~19% of evolution. Contract extraction lifts to ~24%. The remaining ~76% is logic — and overclaiming extraction is a research risk.
Paper 2 extracts 130 clean atomic transitions from 707 SWE-bench GitHub fixes. This article explains how that benchmark is constructed and what its scope allows you to conclude.
If your evolving-knowledge benchmark marks outdated chunks as “[OUTDATED]”, baselines win by reading the label. MemStrata’s marker-free protocol removes that contamination — and scores drop up to 14 points when markers are removed.
Calibration results show cosine similarity cannot separate contradicted facts from duplicated ones. Contradictions are more embedding-similar than rephrases. Here is the data and the design implication.
MemStrata maintains temporal validity with deterministic (subject, relation, object) supersession in a bi-temporal ledger. No similarity threshold. No LLM on the read path. ~2.1s retrieval latency.
When knowledge evolves, the evaluated retrieval baseline retrieves both stale and current facts with near-identical embeddings. Stale-fact errors in that protocol are structural, not a tuning problem.
Install MemStrata locally and see token savings + temporal memory on your own codebase.
Install Conversational