Spoken context becomes useful to an AI workflow only after it is captured, converted into a durable file, given enough provenance to be trusted, and exposed through a bounded retrieval interface. Recording alone does not close that loop. A folder of audio is difficult to search; an unreviewed summary can erase the reason behind a decision; and unrestricted agent access can disclose more than the user intended.

Evidence boundary. This article describes a technical workflow and the documented OneMira C1 data boundary. It does not claim that every transcript is accurate or that local processing makes a workflow legally compliant.

Why does spoken context disappear from AI work?

Agents normally begin with artifacts: source code, issues, design files, project rules, and documents. A desk conversation is absent unless somebody turns it into an artifact. The missing information is often not the final decision but the rejected options, constraints, owners, and uncertainty that explain why the decision was reasonable.

A useful capture system therefore needs to preserve four layers:

  1. Observation: what was actually said, with timestamps or speaker boundaries where available.
  2. Interpretation: a summary, decision, task, or contradiction derived from the transcript.
  3. Provenance: when and where the text came from, plus a link to the source session.
  4. Disposition: whether the information remains raw, becomes durable project knowledge, or is deleted.

If interpretation overwrites observation, later reviewers cannot tell whether a confident sentence was spoken or inferred. Keeping raw notes and derived notes as separate files or clearly separated sections makes correction possible.

What is the minimum capture-to-context pipeline?

The pipeline has two independent quality gates. The first is transcription fidelity: did the text preserve names, numbers, negation, and domain terms? The second is knowledge fidelity: did the derived decision retain its rationale and uncertainty? A fluent summary can pass neither gate.

For C1, the approved product specification places voice activity detection on the device and speech recognition, transcription, and speaker labeling on the paired computer. Encrypted audio slices remain on C1 while capture, transfer, or local transcription is pending; each slice is deleted from C1 after the computer consumes it for transcription. See the current C1 specification for the product-level statement.

OneMira C1 Silver White model rendered from a front 45-degree angle
Exact C1 geometry rendered from the production Three.js model. The visible turntable is a recording-state cue; it does not replace consent.

Why use Markdown instead of a proprietary note database?

Markdown is not automatically better than a database. Its advantage is interoperability. A plain-text note can be inspected with an editor, searched with rg, versioned with Git, opened in Obsidian, and passed to an agent without first exporting from a vendor portal. GitHub Flavored Markdown also has a published syntax specification, so headings, tables, links, and code blocks have predictable semantics.

A practical session file should include stable metadata rather than decorative fields. For example: 

---
kind: session
session_id: s_20260716_093000
created_at: 2026-07-16T09:30:00+08:00
source: c1
review_status: reviewed
---

# Checkout retry discussion

## Decision
Retry idempotent payment-status reads, not payment creation.

## Rationale
Creating a second payment can duplicate an external side effect.

The identifier lets software link derived notes back to the session. The review state tells an agent whether it may treat the text as verified. Dates should use an explicit time zone; filenames alone are a weak source of truth because clocks can be wrong and files can be renamed.

How should an agent retrieve local context safely?

The official MCP architecture separates a host, clients, and servers. A server can expose tools or resources, but the host controls the connection and user interaction. For local servers, the MCP security guidance recommends stdio where practical, minimal filesystem and network privileges, explicit consent, and careful review of startup commands.

That leads to a bounded pattern for voice context:

  • Expose search and read operations before write or delete operations.
  • Constrain the server to the OneMira note root instead of the entire home directory.
  • Return short excerpts and source paths, then retrieve a full note only when needed.
  • Keep third-party model disclosure separate from local retrieval. A local file does not stay local if its text is pasted into a cloud model.
  • Log which local tool ran without logging transcript content or credentials.

The OneMira MCP documentation explains the supported local workflow. The companion Cursor setup and security guide adds a concrete verification sequence.

How do you measure whether the loop works?

Do not measure success by transcript volume. Use an acceptance set drawn from real work and score the points that can change an outcome:

Check Objective pass condition Failure example
Critical-term recall Names, numbers, negation, and identifiers match the reviewed audio “Do not retry” becomes “retry”
Decision provenance Every derived decision links to a source session Rule appears with no traceable origin
Retrieval precision A scoped query returns the relevant note before unrelated sessions Broad keyword floods the context window
Disclosure control The user can identify what text leaves the computer Agent silently sends an entire transcript archive
Recovery An interrupted transfer can resume without duplicate notes or premature deletion Source audio is removed before note commit

Review false positives as carefully as missed items. A system that invents a decision can be more damaging than one that asks for clarification.

What should become durable project knowledge?

Promote information only when it changes future work: an approved constraint, a rejected architecture with a reason, a named owner, a user requirement, or a reproducible failure. Social conversation, repeated status updates, and uncertain guesses normally stay in the session or expire.

A scheduled agent can propose changes to project rules or a wiki, but the proposal should include a diff and source link. C1 is the capture and local-context layer; it does not independently decide which conversation becomes policy.

Sources and further reading