Daily summariser agent: extract facts from conversation and write to session memory

[djthorpe]

Overview

A scheduled summariser agent runs periodically (e.g. hourly or at end-of-day) against the session's conversation history. It extracts structured facts — decisions made, preferences stated, tasks completed, open threads — and writes them into the session memory store (see #41). This gives the assistant a persistent, queryable knowledge layer that survives context window compaction.

This is distinct from context compaction (#19), which manages what fits in the current prompt. The summariser is about building durable knowledge, not just summarising for space.

Behaviour

• Runs as a trigger on a configurable schedule (e.g. hourly, or triggered at end of an active conversation window)
• Reads the conversation history since the last summariser run
• Extracts structured facts using a one-shot agent prompt
• Writes extracted facts as files into the session memory store (e.g. facts/YYYY-MM-DD.json)
• Updates a running facts/index.json manifest so the assistant can query what's known without reading all fact files

Fact schema

Each extracted fact should carry enough structure to be useful:

• type — category (preference, decision, task_completed, open_thread, person, etc.)
• content — the fact itself in plain language
• confidence — high / medium / low
• source_date — when the conversation segment occurred
• extracted_at — when the summariser ran

Integration with session memory (#41)

• Fact files are written into the session filestore under a facts/ prefix
• The assistant can use file_list and file_read tools to retrieve facts on demand
• The index file provides a lightweight lookup without loading all facts
• Semantic search over fact content can be added later as a memory_search tool

Prompt design

The summariser agent prompt should:

• Instruct the model to extract only verifiable, concrete facts (not inferences)
• Avoid duplicating facts already present in the index
• Output structured JSON matching the fact schema above
• Be conservative — better to miss a fact than to hallucinate one

Notes

• The summariser should be idempotent — re-running over the same window should not produce duplicate facts
• Low-confidence facts should still be stored but flagged, so the assistant can treat them with appropriate uncertainty
• This is related to but separate from context compaction (Session compaction: auto-summarize history when approaching context window limit #19) and session naming (Auto-summarize session name from conversation content #18)