feat: use assistant context as meeting memory

2026-07-17 11:26:49 +02:00
parent 407db80413
commit 3cadf08fa2
13 changed files with 188 additions and 2 deletions
@@ -0,0 +1,45 @@
+## Context
+
+Meeting Assistant already creates one assistant context artifact per meeting and gives both the automatic summary agent and the interactive agent window tools that can read and append to it. The current summary prompt calls the file a notebook, while the interactive agent prompt describes only generic artifact repair. Neither prompt establishes assistant context as the durable handoff point for uncertainty, failures, assumptions, repairs, and conclusions about one meeting.
+
+## Goals / Non-Goals
+
+**Goals:**
+
+- Make the assistant context's meeting-specific memory role explicit to both agents.
+- Preserve summarization problems, missing information, and assumptions for later investigation.
+- Make the interactive agent consult that memory before repairing a meeting or summary.
+- Make the interactive agent record completed fixes and conclusions for later agents.
+
+**Non-Goals:**
+
+- Add a new artifact type, schema, section format, or database.
+- Automatically rewrite or summarize existing assistant context files.
+- Change tool permissions or allow agents to access artifacts outside existing configured scopes.
+- Override an explicitly configured custom summary-agent prompt with built-in summary guidance.
+
+## Decisions
+
+### Extend the existing instruction builders
+
+The automatic summary behavior will be added to `MeetingSummaryInstructionBuilder.DefaultInitialPrompt`. This preserves the existing contract that a configured `Agent:InitialPrompt` replaces the built-in summary prompt.
+
+The interactive behavior will be included in the instruction builder's always-appended meeting-artifact guidance. This keeps the meeting-memory reminder available even when the interactive agent has a custom initial prompt, alongside the tool capabilities that Meeting Assistant already appends.
+
+Alternative considered: implement automatic interception or mandatory writes whenever an agent encounters uncertainty. The application cannot reliably infer those semantic events from arbitrary model turns, so explicit instructions are the smallest dependable mechanism.
+
+### Keep memory entries append-oriented and concise
+
+Agents will be told to append problems, missing information, assumptions, fixes, and conclusions to the matching assistant context. Existing `write_context` behavior already appends by default, which preserves earlier observations and avoids introducing a structured migration.
+
+Alternative considered: define mandatory headings or a machine-readable memory schema. That would add formatting and compatibility obligations without being necessary for the requested agent handoff behavior.
+
+### Keep user-facing summary content separate
+
+Assistant context remains internal meeting memory; the generated summary remains the user-facing artifact. Assumptions and missing evidence can inform explicit uncertainty in the summary, but diagnostic notes and repair history belong in assistant context.
+
+## Risks / Trade-offs
+
+- [Repeated agents may append duplicate observations] → Instruct agents to read existing context first and keep entries concise.
+- [Assistant context can grow over time] → Preserve existing ranged and tail reads; no new retention mechanism is introduced.
+- [Configured custom summary prompts omit the new built-in guidance] → Preserve the documented replacement semantics and leave responsibility with the custom prompt author.