Omni Memory — Joint Audio-Visual Memory for Long Videos

Check the qwen-mm-plugins-omni-memory tools in your tool list for full schemas and parameters.

First decide whether this video needs a memory at all

Building one costs an omni call per 30s window, so over a short video it spends N calls indexing what fits in a single request. Decide by length, and by whether the user wants a memory.

How you learn the length: get_memory_status(video_path=...). It reports duration_min for the source video even when no memory exists, and its next_step already applies the table below — so the call you had to make anyway also answers this. (It reads the file with ffprobe; if that is unavailable the field is absent and building is the safe default.)

Two things that override the table:

• Several questions about the same video → build the memory, even a short one. A watch is stateless: every question re-uploads and re-watches, while a memory is paid for once and then answers for free. One question is cheaper as a watch; a conversation is cheaper as a memory.
• A memory already exists → use it. watch_and_answer re-uploads the video and sees only what is in it; the memory already holds every utterance with its speaker, and answers in milliseconds.

If a watch cannot get through, its result says so explicitly: fallback: "build_memory" plus the exact command. Run that and continue with the flow below. When the failure is throttling or a misconfigured endpoint there is no fallback field — retry or fix the configuration instead, because a build would hit the very same wall.

General Workflow

Everything below is the memory path. Memory must exist before you can query it. When a user asks about an audio-video, you are the one who retrieves memory and answers — every tool here returns evidence, not an answer.

1. get_memory_status(video_path=...) — always start here. It reports one of:
   • exists: false → also carries duration_min, and a next_step that applies the routing table above: short enough to watch instead, or build it (see "Building memory")
   • exists: true, complete: true → query it
   • exists: true, complete: false → truncated. Its answers cannot be trusted. An interrupted build still finalizes the library, so it looks normal otherwise; continue it with --mode resume
2. Query it (see "Retrieval workflow")
3. Answer from what you retrieved. The record already carries the audio's content — every utterance with its speaker, how it was said, and the non-speech sound around it — so this is where most questions end.

replay_and_answer is not step 4. It is an exception off the side of this flow, taken only when you can name a specific thing the record does not contain — and it re-watches only the few clips you name, so on everything else it knows less about the video than the memory does.

The Memory Layout

Four retrievable containers plus the clip files, filled by one pass over the video:

Entities     canonical people: person_id, resolved name, appearance, attributes
Semantic     induced facts as keyed triples, e.g. David/role → "…"      ← by EXACT key
Episodic     one record per 30s window: visual caption + every utterance
             with its speaker_id, paralinguistics, acoustic events        ← by hybrid search
Scene env    durable environment/layout items, recalled on demand
Clips        the 30s files themselves, so an answer can re-watch the source

Three properties drive how you retrieve:

• person_id (P001, P002, …) is stable for the whole video. Names are inferred separately by reading the accumulated transcript, so a person may be P003 with name: null until someone addresses them by name — expected, not a failure. For those people, get_memory_overview carries also_heard_as (the names they were heard called, e.g. P003 also_heard_as ["Dara"]); that is how you map a name the user says onto an anonymous id.
• Semantic keys are an exact lookup, not a search. Keys are built from the resolved name (David/prefers/transport). get_memory_overview gives you the complete directory, so you pick from it rather than guessing — a key that does not exist simply returns nothing. Note which way round this goes: the container named Semantic is the one reached by key, while hybrid search is what reaches Episodic. To find a fact by its content instead of its key, use search_facts(query=…) — a separate tool, not something a plan can do.
• Environment items are on demand, not resident. Recalling them for every question measured net-negative, so include_scene is a deliberate pick for "where is X" questions.

Building memory

build_memory/ runs through Bash with the system Python, outside the MCP server's uvx environment. It requires Python 3.10+, pip, ffmpeg/ffprobe, and DASHSCOPE_API_KEY; missing Python packages (numpy<3, openai) are installed automatically. Run qwen-mm-plugins-omni-memory --check-system before building.

# one video → memory next to it, at <video>.memory/
python3 script/build_memory/build_memory.py /path/to/video.mp4 --model qwen3.5-omni-plus

# many videos → independent per-video memories, in parallel
python3 script/build_memory/build_memory.py --video-dir /path/to/dir -j 4 --model qwen3.5-omni-plus

python3 script/build_memory/build_memory.py /path/to/video.mp4 --mode rebuild  # discard and start over

--model is the omni model all three build stages use — per-clip extraction, semantic induction and name alignment. It defaults to qwen3.5-omni-plus (or $QWEN_MM_API_OMNI_MODEL if set), so it can be omitted; name it explicitly when you want the record to say which model produced the memory. The endpoint comes from DASHSCOPE_BASE_URL, or DashScope's default host when it is unset.

The script writes its own log next to the memory (build_<timestamp>.log, overridable with --log) and prints the model, endpoint and log path on startup, so there is nothing to redirect.

Paths above are relative to this Skill directory. Everything the build needs is inside script/build_memory/ — it imports nothing from the MCP server's package, so it runs the same whether that package is installed or not.

Streaming several videos into one memory

Give the same --namespace to several videos and they become one continuous memory:

python3 script/build_memory/build_memory.py --video-dir /path/to/session --namespace my_stream
python3 script/build_memory/build_memory.py next_part.mp4 --namespace my_stream --mode append

Each segment starts from the previous segment's state, so a person keeps the same person_id across videos, semantic facts keep accumulating and merging, and timestamps are stitched end to end into one timeline. Use it when several files are really one recording session. Segments must be given in chronological order, run strictly serially (-j is ignored and reported), and a failed segment stops the stream — later ones would start from missing state.

By default it lives at <first-video-directory>/<namespace>/; query it with both video_path and namespace. Set MEM_LOCAL_DIR when you want a fixed shared library and namespace-only queries. Videos appended in separate commands must share a directory unless MEM_LOCAL_DIR is configured.

Requirements: ffmpeg/ffprobe on PATH and DASHSCOPE_API_KEY. Omni calls use DASHSCOPE_BASE_URL and QWEN_MM_API_OMNI_MODEL when configured.

Retrieval workflow

Step 1: pick an entry point by question type

Targeted questions go straight to their tool — do not route everything through plan_and_search.

Step 2: orient, plan, search, answer

For the open-ended case, you are the planner. plan_and_search executes the plan you give it and never answers. One search per question:

① orient — get_memory_overview
   → people (person_id, name, appearance, also_heard_as)
   → semantic_key_directory   ← the complete list of fact keys; you pick from it, never guess
   → scene_env_available      ← a marker only; ask for the items via include_scene below

② plan it yourself, then ONE search
   plan_and_search(question="What did they agree on in the end?",
                   people=["P001","P002"],                    # from step ①
                   fact_keys=["event:campaign/decision"],      # from the key directory
                   queries=["two people settle on a plan",     # descriptive statements
                            "someone agrees to a proposal"],
                   time_ranges=[[1500, 1800]],                 # only if time-bounded
                   include_scene=False)                        # on only for "where is X"

③ read the evidence and answer.
   ↑ for most questions the flow ends here

   ⤷ exception, only if you can name the specific unrecorded detail you need:
     replay_and_answer(idxs=suggested_replay_idxs, question=...)

Step ① is not optional and not a probe: fact_keys is an exact lookup, so a plan that names no keys gets no facts at all. The directory is what makes the plan possible — get it first, then plan once. It stays in your context for the rest of the conversation, so one call covers every later question about the same video.

If a search comes back thin, re-plan and search again — but change the axis (name people, name keys, add a time range) rather than rewording the same queries.

Decision flow

Question arrives
  ├─ Short video (<10 min), or user wants no memory (and it is <30 min)?
  │                                          → watch_and_answer — ends here, no memory involved
  ├─ get_memory_status                       ← never skip; truncated libraries look fine
  ├─ Names one person, asks what they said?  → get_person_dialogue
  ├─ Asks who said something?                → search_dialogue
  ├─ Asks about a fact / role / relation?    → search_facts
  ├─ Names a time?                           → get_timeline → get_moment
  └─ Open-ended?                             → get_memory_overview → plan → plan_and_search → answer
                                                                                              ↑ ends here

replay_and_answer is not on this flow — see Usage Rules

Tools Reference

Status and orientation

get_memory_status — Does the memory exist, is it complete, and does this video need one?

• Params: video_path or namespace
• Returns: exists, complete, clips/episodic/planned_clips, people, named_people, semantic_facts, duration_sec; truncated and next_step when incomplete
• When exists: false it also reports the SOURCE video's duration_sec / duration_min and a next_step that applies the routing table at the top — this is where you learn whether the video is short enough to answer with watch_and_answer instead of building anything
• Use when: always, first. complete: false means the answers are not trustworthy.

get_memory_overview — The vocabulary you plan with. Read this before plan_and_search.

• Params: video_path or namespace
• Returns: people (person_id, name, appearance, and also_heard_as for anyone still unnamed), semantic_key_directory (complete), scene_env_available
• Use when: before planning any open-ended retrieval. It costs one call and stays in your context for every later question about the same video.
• scene_env_available is a marker, not the items: pass include_scene=True to plan_and_search when the question needs them

Answering an open-ended question

plan_and_search — ONE retrieval from a plan you decide, fusing the containers. Does not answer.

• Params: question, people, fact_keys, queries, time_ranges, include_scene (off by default — turn it on for "where is X"), top_k (default 5)
• Returns: people, facts, moments (briefs), scene_env, evidence_text, suggested_replay_idxs, plan_used
• fact_keys is an EXACT lookup — pick from get_memory_overview's semantic_key_directory. Naming no keys returns no facts; that is why the overview comes first
• Query guide: queries take DESCRIPTIVE STATEMENTS, not questions
  • Good: "someone offers to bring an umbrella"
  • Bad: "who brought an umbrella?"
  • Several short angles beat one long query — each contributes its own top_k recall
  • These also decide suggested_replay_idxs, so describe what to look for

replay_and_answer — Re-watch clips with their audio and have the omni model report what it sees.

• Params: question, idxs, evidence (optional text context), model (optional omni model; defaults to qwen3.5-omni-plus, and need not match the model the memory was built with)
• Returns: answer, watched_idxs, model, sent_mb; dropped_idxs / missing_idxs when applicable
• It answers instead of returning evidence, because what it reads is in the video itself
• Watches at most 3 clips per call (~3.5 MB each, inline) — every other clip stays invisible to it. Extras come back in dropped_idxs
• Use when: one specific detail the 30s record demonstrably lost — a fleeting object, a facial expression, a count
• Not for what was said. Every utterance is stored verbatim with its speaker and paralinguistics, so get_person_dialogue / search_dialogue / get_moment answer wording questions better, and for free

Answering without a memory

watch_and_answer — Watch a SHORT video whole, in one call, with no memory involved.

• Params: video_path (required — a source video, NOT a namespace), question, model (optional)
• Returns: answer, sent_mb, duration_sec, transcode_cached, model
• On failure: error + failure (reject / timeout / empty / rate / config). For the first three it also carries fallback: "build_memory" and next_step — run that command, then use the memory path. For rate (throttling) and config (wrong endpoint or key) there is deliberately nofallback: a build runs against the same endpoint and would fail identically
• Use when: the video is under ~10 min, or the user has said they do not want a memory built. See the routing table at the top — over ~30 min this is not an option
• The video is re-encoded once and that copy kept, so follow-ups skip the re-encode (transcode_cached: true). The omni call is still paid every time, which is why several questions are cheaper as a memory
• It answers rather than returning evidence for the same reason replay_and_answer does — what it reads is in the audio, which you cannot hear — but it sits off the retrieval flow, before there is anything to retrieve

Direct access

get_people — Full dossier for everyone, or one person_id.

• Params: person_id (optional)
• Returns: person_id, name, appearance, attributes, first/last appearance, clip indices
• Use when: "who are these people", or you need one person's full record

get_person_dialogue — Every line one person spoke, in order.

• Params: person_id (optional — omit it for every speaker in time order), start_sec, end_sec, limit
• Returns: total and the utterances with timestamps and paralinguistics
• Use when: "what did X say" — this is exhaustive, where plan_and_search is top-k

get_timeline — Moments in a time range, chronologically.

• Params: start_sec, end_sec (end_sec=0 means "to the end")
• Returns: brief per moment — follow up with get_moment for detail
• Use when: the question names a time and there is nothing to search for

get_moment — Full detail for specific clips. Most information-dense tool.

• Params: idxs (required)
• Returns: full visual_caption, scene_continuity, scene_env_update, per-person actions, every utterance, acoustic_events, and clip_path
• Use when: a brief is not enough. clip_path is the 30s file on disk, should you have another tool that can open it

search_dialogue — Semantic search at the UTTERANCE level; every line comes back with its speaker.

• Params: query (required), top_k
• Use when: "who said …" — you know the content but not the person

search_facts — Three ways into the semantic container.

• Params: one of query (search statements by content), key_prefix ("David/", or a person_id prefix, which is mapped for you), subject_id; plus top_k
• Use when: you want facts by CONTENT rather than by exact key — plan_and_search cannot do this

search_memory — Broad hybrid search across containers, with entity-anchored multi-hop.

• Params: query (required), top_k
• Use when: you know nothing about the library and want one wide net. plan_and_search is usually the better entry point.

Usage Rules

1. On the memory path, never answer without checking status first — a truncated library answers confidently from half a video, and looks entirely normal while doing it. (watch_and_answer is not on that path and needs no status check; it reads the video, not a library.)
2. The text memory is the answer path; replay is the exception. Before reaching for it, you must be able to name the specific thing you need that the record does not have. "The record is only a summary" is not such a reason — a replay sees only the clips you name and nothing else, so a vague replay returns a vaguer answer than the memory you already hold.
3. Stop when the open-ended path and a replay have both failed. get_memory_overview → plan_and_search → replay_and_answer is the deepest this memory goes; there is no further tool that knows more. Do not go fishing through the others — say what the evidence did establish and what it could not, and leave it there. More calls at that point produce a guess, not an answer.