Introducing Temporal Validity Memory: How MemStrata Retires Stale Facts Without an LLM

MemStrata is a local retrieval memory between your AI coding tool and your LLM provider. Like RAG, it stores facts extracted from interaction. Unlike RAG, when a fact’s value is contradicted, a deterministic supersession rule retires the stale value in a bi-temporal ledger — with no similarity threshold and no LLM call on the critical path.

This is the core of our paper Temporal Validity in Retrieval Memory (arXiv:2606.26511): eliminate stale-fact errors for agents over evolving knowledge while preserving static recall. That research mechanism is not the same as the product. Conversational keeps the index on your computer. You connect your own AI.

Each incoming turn yields a candidate fact:

1. Exact-duplicate short-circuit — normalized text hash drops verbatim repeats.
2. Deterministic assertion path — clean (subject, relation, object) triples normalize the (subject, relation) key. Different object → supersede. Same object → reinforce. No prior key → novel store.
3. Text-gate fallback — non-triple prose uses similarity + judge only as a secondary path; non-contradictory near-duplicates are retained as distinct facts.

An early “lossy” variant that aggressively merged near-duplicates regressed below RAG on static recall. The published design therefore retains like RAG and bounds growth only on the axis that matters: contradictions.

Query embedding → top-k over active facts → pack original source sentences (not terse reconstructed triples). No language model runs on the read path, so latency sits at the embedding floor (~2.1s) versus ~16–18s for LLM-reranking and verification baselines on the same local stack.

All experiments in Paper 1 run entirely on consumer hardware with Qwen2.5-Coder-7B, temperature 0, fixed seeds, and no network — so the claims are reproducible, not vendor-demo cherry picks.

For developers using Cursor, Claude Code, VS Code, or Cline, every confidently wrong answer about “what the port is now” burns tokens and trust. Temporal validity is not a nicer dashboard metric — it is the difference between an assistant that compounds knowledge and one that gaslights you with last month’s truth.

Install the product that matches the job. Coder supplies selected repository context to a supported coding tool. Conversational does not ship an answering model. You connect the AI you already use. Indexes stay on the device; account services and any host or cloud model you connect can send data off the device. See Trust.