💡 Voice-Searchable Session Archive: Conversational Retrieval of Past Decisions and Outputs

[github-actions[bot]]

Summary

Users ask the avatar to find past work by speaking naturally — "find the session where we discussed the pricing model" or "what did we decide about the auth migration?" The avatar searches session summaries and decision metadata (FR46-47), presents the top matches as ActionCards, and can synthesize answers across multiple sessions. This transforms TalkTerm's session history from a passive log into an active, voice-searchable knowledge base that makes accumulated work findable and reusable.

Market Signal

ChatGPT offers basic text-based conversation history search but no voice-first search with avatar-narrated results. Enterprise knowledge management is a $25B+ market (Gartner 2026), and the gap between "information exists" and "information is findable" is consistently cited as a top knowledge worker frustration. No AI desktop agent offers voice-searchable session archives with conversational result presentation. The closest competitor feature — ChatGPT's history search — is text-only, returns raw conversation snippets, and offers no synthesis across sessions.

User Signal

The PRD mandates cross-session memory (FR46-47) and session resume (FR31), but current retrieval is limited to "most recent incomplete session" (FR33). As users accumulate sessions, finding specific past decisions becomes critical — a PM who brainstormed pricing models three weeks ago shouldn't need to scroll through 20+ sessions to find that decision. The PRD's success metric of daily usage (≥1 session/day) means 30+ sessions in the first month alone. Without search, past work becomes invisible, and users lose confidence that TalkTerm remembers.

Technical Opportunity

SQLite already stores session data per the architecture (better-sqlite3). SQLite FTS5 provides full-text search with minimal overhead — well-suited for the structured session summaries mandated by FR46-47. Voice queries are natural language, parsed by Claude (the existing agent backend) into search parameters. Search results are displayed as ActionCards (UX-DR2) in the existing left panel, with session preview in the OutputPanel (FR45 document mode). Architecture is fully aligned: this is an incremental extension of existing data and UI components, not a new system. Implementation touches the storage bounded context (main/storage/) for FTS5 index, agent bounded context (main/agent/) for query parsing, and renderer (components/) for result display.

Assessment

Dimension	Score	Rationale
Feasibility	high	SQLite FTS5, existing session data, existing ActionCard/OutputPanel UI — incremental, not greenfield
Impact	high	Critical for knowledge continuity; transforms passive logs into active knowledge base; unique voice-first UX
Urgency	med	Value grows with session count — becomes critical after first month of daily use

Adversarial Review

Strongest objection: Search quality depends entirely on the quality of session summaries (FR46-47). If summaries are poor or inconsistent, search results will be unreliable and erode user trust.

Rebuttal: FR46-47 already mandates structured summaries with decisions, project vocabulary, and workflow outcomes — the summary quality investment is required regardless of search. FTS5 supports relevance ranking, and Claude can augment keyword search with semantic understanding of voice queries ("the one about pricing" → search for "pricing", "cost model", "revenue"). Start with exact-match search on session titles and decision keywords; expand to semantic search as the summary corpus grows. The avatar honestly acknowledges search uncertainty: "I found 3 sessions that might be relevant — here are the closest matches" rather than presenting uncertain results as definitive.

Suggested Next Step

Define the FTS5 schema extension for session summaries (indexed columns: session title, decision keywords, output artifact names, workflow type). Design the voice query → search parameter mapping via Claude. Create a UX spec for search results as ActionCards with session preview in the OutputPanel, including "no results" and "multiple matches" conversation flows.