---
source_hash: 2e6b2237e9d953790f78c4387fc10320995a5e3e7ae4b5a890c2db5cefe08510
post: /posts/tools-and-plugins/odvm-video-memory
---

# Context Pack: ODVM video memory

Reference for the rewritten August 2026 ODVM v0.1.0-rc1 account. Product route: /documents/tools-and-plugins/odvm-video-memory. This is a dated architecture/release account, not a newly audited package. No installation procedure is established here.

## Verbatim excerpts

> Sidecars are the retained observations.

> No vision re-run does not mean no model work or a guaranteed price per query.

> The model is replaceable. The memory you build is the asset.

## Four stages

Sample: cuts anchor keyframes, motion increases coverage, and a minimum sampling floor covers static stretches. Samples can miss events; more samples do not guarantee correct descriptions.

Observe: the described local vision model writes observations in versioned sidecars keyed by model. Local speech writes utterances with start/end times on the same timeline. Ordinary queries reuse observations; a deliberate model change can create a new observation version alongside the old. Do not say a clip can only ever be observed once.

Index: SQLite full-text search plus embeddings. The intended rebuild contract uses retained sidecars and the information needed to interpret them; the index is regenerable. Model observations are not guaranteed truth. Footage remains the evidence for important claims.

Query: returns ranked clip/time spans and matching observations without re-running vision. Semantic query embeddings can still require inference. `odvm pack` is the described path to a sized agent context bundle with timestamps. No constant-cost or zero-inference promise.

## Original command example

```bash
odvm sample  --project myproject --asset clip.mp4 --job edit_shoot
odvm observe --project myproject
odvm index   --project myproject
odvm query   --project myproject "the red jacket walks left"
```

Two profiles share schema/pipeline: edit_shoot favours visual ranking for takes/plates; stream_personal favours transcripts for long recordings while retaining silent visual events. The release account includes gold reference queries to compare recall after prompt/rank changes, not a universal quality certificate.

## Availability and teaching surfaces

Catalogue: content/downloads/odvm-rc.mdx, tier ARCHITECT and above. Route: /app/api/downloads/odvm-rc. Asset key: downloads/architect/odvm-rc/odvm-0.1.0-rc1.zip. The private package was not verified in this rebuild; do not invent `pip install` instructions from the older pack. Local processing is the reported August setup, not proof of arbitrary configurations.

Five retained PNG diagrams explain the pipeline, speech/visual coverage, sampling, index and query. The old sidecars-are-truth and query-cost language is qualified in captions/narration. The 0.93/0.87/0.81 figure values are illustrative.

DensitySliderFidget: synthetic 60-second timeline, six fixed cuts, hand-set motion/static regions, density 0–100, step 5, default 40, floor interval 6 seconds and 1–8 motion samples per region. QueryPlaygroundFidget: seven hand-written moments, keyword overlap after stopword removal, top three matching spans, earlier time breaks ties. Percentages are overlap, not confidence or observed model output. Both live under apps/web/components/blocks and make no video/model calls.

Keep the sibling link /documents/tools-and-plugins/lens, the pullquote, all five diagrams and exact CLI block.