Plain-language summary. Several things the Companion should do all need the same underlying capability: understanding what a note MEANS, well enough to relate it to other notes — not by matching keywords, but by meaning. This foundation builds that capability (embedding-based semantic understanding) on top of the trustworthy memory Room 1 now guarantees (never-forget, corrections-preserved, provenance, actor-boundaries, write- and personal-recall-truthfulness).
The industry chases better retrieval/context on shaky memory. Loomworks builds meaning-understanding on memory that can't lie. That is the differentiator — and the foundation being cemented.
The 2026-07-01 testing showed THREE capabilities that all need meaning-understanding — they are siblings, not separate problems:
ask_about_past_input today retrieves ALL committed notes and lets the model focus on the topic (no filter). It recalls TRUE facts (verified: "where am I parked" → "level 12", which is genuinely the most recent recorded parking fact — NOT a fabrication), but it's model-composed and unfiltered. The foundation makes it: retrieve notes by MEANING (find what's actually about X), then server-compose a faithful readout of exactly the retrieved set (the CR-129 truthful-readout pattern, applied to a meaning-filtered set). Focused AND faithful.
The parking facts (Level 8, Level 8 again, 13, 12) are FOUR separate assertions the system doesn't know are about the same thing (where the Operator parks). Same disease as the Shane-birthday duplicate and the meeting-note duplicates (R2). When the Operator says "I'm parked on 12", the system should recognize this is a new value for an existing fact and supersede (creating the 8→13→12 trail, corrections-preserved) rather than capture a fourth loose note. Recognizing "this updates fact Y" needs the same meaning-understanding as "this is about X". So correction-recognition folds into this foundation (Operator decision 2026-07-01), it is not a separate R2 patch.
Mechanism when recognized: SUPERSEDE (keep the old value as history, marked corrected, recall surfaces current) — NOT retract. Parking moved over time; that's a correction with real history, not noise. Retract is for wrong-facts-to-erase; supersede is for out-of-date-facts.
The routing question (does a note belong in THIS engagement or elsewhere) is also a meaning judgment — "is this note about Fenwick, or about a work meeting?" Likely draws on the same substrate. Not necessarily in the first build, but recorded as related so it's not re-derived later.
The insight (2026-07-01): meaning-understanding is not a recall feature — it is the substrate under recall, correction-recognition, and routing. Cementing it pays off in all three. That is why it is THE first foundation-effort.
show_held/recall_personal mechanism) so the meaning-filtered recall is truthful-by-construction.FIRST effort after Room 1 closes (Operator directive: cement foundations). Grounded and built deliberately — not rushed. Precedes the other queued work (upload pathway, held-tray robustness, wide-net reduction) because it is the substrate several of them lean on.
Note the Operator-authority thread runs through it: correction-recognition should PROPOSE supersession and let the Operator confirm (machine surfaces, Operator approves) — consistent with the whole Memory-room design. Semantic capability finds and proposes; it never silently rewrites memory.