mneme

# Mneme Use this skill when processing conversational transcripts, denoising audio notes, synthesizing action items and summaries from discussions, or querying historical meeting context. ## Overview Mneme structures meeting and conversational intelligence into three distinct phases: 1. **Lossless Denoising (Zero Information Loss)**: Clean raw ASR audio transcripts, eliminate speech recognition loops and verbal tics, and reconstruct accurate speaker turns without summarizing or dropping facts, numbers, dates, decisions, or tangents. 2. **Synthesis and Action Extraction**: Generate structured executive summaries, key decisions, and actionable task items (with clear context, scope, and completion criteria) suitable for any task management system or personal checklist. 3. **Contextual Retrieval (Talk with your notes)**: Search and retrieve exact historical dialogue and agreements across indexed notes using local hybrid search. ## Public entrypoint ```bash uv run --script skills/current/mneme/scripts/cli.py <command> [options] ``` ## Required follow-up reads | Need | Read | When | | :--- | :--- | :--- | | Lossless transcript cleaning rules | `references/lossless-cleaning.md` | Denoising raw transcripts without dropping information | | Action item & summary schemas | `references/spec-synthesis.md` | Converting conversations into structured tasks and summaries | | Search engine integration reference | `references/qmd-integration.md` | Searching, indexing, or querying meeting archives | ## Common calls ### 1. Denoise a raw transcript losslessly ```bash uv run --script skills/current/mneme/scripts/cli.py denoise raw_transcript.md -o clean_transcript.md ``` ### 2. Synthesize summary and actionable tasks ```bash uv run --script skills/current/mneme/scripts/cli.py synthesize clean_transcript.md --summary summary.md --tasks tasks.json ``` ### 3. Hybrid search over indexed meeting knowledge base ```bash uv run --script skills/current/mneme/scripts/cli.py search "quarterly budget timeline" --json ``` ### 4. Search exact terms via keyword search ```bash uv run --script skills/current/mneme/scripts/cli.py search "Project Delta" --exact ``` ### 5. Fetch a line-ranged transcript snippet ```bash uv run --script skills/current/mneme/scripts/cli.py get "#<docid>:<line>:<count>" ``` ## Core Rules - **Lossless Invariance**: When cleaning transcripts, NEVER summarize, drop, or condense dialogue, metrics, dates, or tangents. Only remove verbal tics, stutter loops, and false starts. - **Strict Diarization**: Preserve every speaker change in chronological sequence. NEVER merge alternating dialogue turns into a single speaker block or insert artificial topic headings into a transcript. - **Actionable Synthesis**: Ground participant names against any available roster. Convert agreements into unambiguous tasks with clear owners, concrete scope, and explicit completion criteria. - **Targeted Retrieval**: Fetch exact line-ranged snippets (`#<docid>:<line>:<count>`) instead of reading whole transcript files into the context window.