Decision snapshot

Status: No deployable K × Q winner as of 2026-07-15.

The current operating architecture remains deterministic ingest plus governed compile workflows for writes, metadata-v2 Markdown as durable memory, and D1 policy-aware sparse/graph retrieval for reads.

  • Query-only formal winner: F1 adaptive router.
  • Operational read default: D1 policy/trust rerank.
  • Coupled diagnostic leader: K3 policy incremental × Q2 policy graph, 31/80 end-to-end cases.
  • Deployment decision: K3 × Q2 is the next architecture target, not a production winner.

What this Lab contains

The child reports record the separate K compilation, Q conditional retrieval, and K × Q coupled evidence; context and latency costs; scaling and maintainability observations; privacy and trust gates; MCP and Agent Skill status; and the next experiment boundary.

Evidence boundary

No relevant internal Notion page was found in the workspace search, so no unrelated Notion content was used as evidence. The authoritative local provenance remains in the LLM Wiki raw research ledger and frozen experiment artifacts. No real private-personal content, prompt bodies, canaries, or credentials are copied into this page.

The local paths cited in the original report (wiki/syntheses/personal-llm-memory-architecture-evaluation.md and related raw ledgers) were not present in the llm-wiki tree at migration time. The Notion text is kept here as the remaining copy.