[Memory P3] LLM-assisted session-end fact extraction #250

Open
opened 2026-04-09 00:41:37 +00:00 by Rockachopa · 0 comments
Owner

Context

The holographic provider has session-end auto-extraction, but it's regex-only (matches "I prefer X", "we decided to Y"). This misses most extractable knowledge.

Task

  1. At session end, use a small local model (gemma4) to extract key facts
  2. Prompt: "Extract durable facts from this conversation. Only facts that will matter in future sessions."
  3. Store extracted facts in holographic fact store with category and trust 0.5
  4. Gate by session length (skip <5 turns) and content (skip pure code/debug sessions)
  5. Deduplicate against existing facts

Design Constraints

  • Must use local model only (sovereignty)
  • Must not add >5s to session teardown
  • Extract 3-5 facts max per session (quality over quantity)
  • Trust starts at 0.5 — rises with use, decays without

Acceptance Criteria

  • After a meaningful conversation, 3-5 relevant facts auto-extracted
  • Facts are properly categorized and entity-tagged
  • No duplicate facts created
  • Works with gemma4:latest (local Ollama)

Part of: [EPIC] Unified Memory Architecture

## Context The holographic provider has session-end auto-extraction, but it's regex-only (matches "I prefer X", "we decided to Y"). This misses most extractable knowledge. ## Task 1. At session end, use a small local model (gemma4) to extract key facts 2. Prompt: "Extract durable facts from this conversation. Only facts that will matter in future sessions." 3. Store extracted facts in holographic fact store with category and trust 0.5 4. Gate by session length (skip <5 turns) and content (skip pure code/debug sessions) 5. Deduplicate against existing facts ## Design Constraints - Must use local model only (sovereignty) - Must not add >5s to session teardown - Extract 3-5 facts max per session (quality over quantity) - Trust starts at 0.5 — rises with use, decays without ## Acceptance Criteria - After a meaningful conversation, 3-5 relevant facts auto-extracted - Facts are properly categorized and entity-tagged - No duplicate facts created - Works with gemma4:latest (local Ollama) Part of: [EPIC] Unified Memory Architecture
Rockachopa added this to the Unified Memory Architecture milestone 2026-04-09 00:41:38 +00:00
Rockachopa added the phase:3-episodicepic:memory labels 2026-04-09 00:41:38 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Timmy_Foundation/hermes-agent#250