hermes - 💡(How to fix) Fix Feature: UnifiedContextEngine plugin — coordinated MemGPT + BigMemory + Cognitive memory [1 participants]

Feature: UnifiedContextEngine Plugin

Motivation

The current Hermes Agent has several context/memory subsystems that are loosely integrated:

1. ContextCompressor (agent/context_compressor.py) — LLM-based message compaction
2. Big-Memory — Snapshot before compaction for recovery
3. MemGPT-style — Hierarchical memory with core/episodic/semantic/archival
4. Cognitive-Memory — Reflection engine, fact extraction
5. Hermes Memory Tiering — LivingMemoryLoader for tiered core memory

These are valuable but independently operated. A unified engine can coordinate them.

What UnifiedContextEngine Does

plugins/context_engine/unified/__init__.py implements a single plugin that:

1. Coordinates all subsystems: Rather than 5 independent systems, one engine orchestrates them
2. Extracts and persists facts: On each compress, calls remember_fact(), add_entity(), add_lesson() automatically
3. Uses SQLite for facts/entities: Maintains a memories.db for cross-session facts
4. Unified snapshot: Takes a single snapshot before any compaction
5. Preserve-on-failure: If any subsystem fails, the original messages are preserved (not partially committed)

Key Design Decisions

• SQLite facts DB: Cross-session persistence of extracted facts, entities, lessons
• Protected ranges: _PROTECT_FIRST_N (default 3) + _PROTECT_LAST_N (default 20) turns never compressed
• Threshold: 80% of context length triggers compression
• Auto-extract: Every compress() call extracts new facts/entities/lessons and persists them

Integration Points

# config.yaml
context:
  engine: "unified"  # activates this plugin

Open Questions for Upstream

1. Should facts/entities be stored in the existing memory/ directory or a new location?
2. Should the extract-and-persist happen at compress time (costly) or async/batched?
3. What's the eviction policy for the facts DB?