- affect.observed and affect.summary exist in the memory event schema.
- affect_frames links back to memory_events by event id.
- Unverified rows keep reduced payloads; full embeddings stay behind speaker verification.
Case Study 04
Snapshot: May 2026
Perception Stack, affect events with runtime boundaries.
The current proof is the MemOS affect lane: typed affect events, an affect_frames projection table, V/A/D versioning, replay and backfill scripts, and read surfaces that only expose verified frames for adaptive context. The broader multimodal vision stays clearly marked as roadmap work.
Proof Surfaces
The public proof is schema, projection, read contract, and boundary.
The old page described future sensing work as if it were already deployed. This rebuild keeps the durable MemOS affect infrastructure visible while separating active, pending, and planned pieces.
Current Build
This is governed affect infrastructure, not magic awareness.
The durable pieces are the event contract, projection table, V/A/D mapping, read helpers, and tests. The page avoids implying unsupervised emotion writes, live video inference, or private affect exposure.
- The projection applier derives missing V/A/D values from the versioned mapping.
- Backfill is idempotent and guarded by typed confirmation.
- Replay only touches the affect projection lane and checks row-count parity.
- Current emotion and history queries filter to verified frames.
- Stale or unavailable data returns fallback payloads instead of pretending certainty.
- Dashboard payloads summarize verification mix, lag, drift, and distributions.
Current Vs Planned
The roadmap stays visible because perception claims drift fast.
- Schema, migration, projection applier, V/A/D parity checks, read helpers, and dashboard payload shaping.
- Targeted local test run passed with 11 passing tests and 3 expected pending gates.
- Public proof uses diagrams and source summaries, not private logs or raw audio.
- Producer repair so live emotion rows stop landing as unknown.
- Non-blocking affect emitter and projection-lag integration checks.
- Speaker verification enforcement before full affect writes are promoted.
- Dedicated pull-tool surface after the memory read contract stabilizes.
- Broader multimodal capture, visual cues, and cross-modal disagreement signals.
- Private calibration profiles and runtime intervention policy after measurement proves useful.
Safety Boundary
Perception enters reasoning only after trust checks and fallbacks.
- Schema shape, projection flow, read helper behavior, and dashboard payload categories.
- Versioning and replay controls that make affect data auditable.
- Current limits, expected pending tests, and roadmap boundaries.
- Raw audio, private transcripts, speaker embeddings, model outputs tied to a person, and live endpoints.
- Unverified frames as a reasoning source.
- Operational hostnames, tokens, connection strings, and private memory contents.
Stack
MemOS events, Postgres projections, pgvector, V/A/D mapping, and verified read helpers.
- MemOS
- Postgres
- pgvector
- affect_frames
- projection_state
- V/A/D v1
- Verified reads
- Fallback payloads
- Replay scripts
- Backfill scripts
- Dashboard stream
- Roadmap boundary
Need perception signals that can survive audit?
That means schemas, trust gates, projection replay, fallback behavior, and clear boundaries before the model ever adapts to a signal.