feat(memory): add enhanced-memory plugin — two-tier store with condensation and semantic search

[labaznov]

Summary

A new memory provider plugin (enhanced-memory) that offers a structured, two-tier architecture for persistent agent memory.

Features

• Two-tier fact storage: raw_facts → condenser → condensed — raw conversational facts are periodically grouped, deduplicated, and summarized into priority-ranked condensed entries
• FTS5 full-text search on both tiers with auto-sync triggers
• Pluggable embedding providers for semantic vector search:
  • Gemini API (default, 3072-dim)
  • OpenAI API (1536/3072-dim, supports custom base_url for Azure/vLLM/Ollama)
  • Local sentence-transformers (no API needed, cpu/cuda/mps)
• Automatic fact extraction from conversations (EN + RU patterns)
• Priority-based condensation with category-aware scoring (security=9-10, user_pref=8-9, etc.)
• Thread-safe SQLite with WAL mode
• Hooks: on_session_end, on_memory_write, on_pre_compress

Architecture

Dialog → auto_extract → raw_facts (SQLite + FTS5)
                              ↓ condense
                         condensed (7 categories, priority 1-10)
                              ↓ embed (Gemini / OpenAI / Local)
                         vec_memory (sqlite-vec KNN search)

Configuration

memory:
  provider: enhanced-memory

plugins:
  enhanced-memory:
    db_path: $HERMES_HOME/memory_store.db
    auto_extract: true
    auto_condense: true
    embedding_provider: gemini  # or openai, local, none

Categories

Category	Priority	Description
security	9-10	SSH, keys, passwords
user_pref	8-9	User preferences
decision	7-9	Decisions and choices
project	7	Project details
tool	6-8	Tools and configs
env	5	Environment/infra
general	4	Everything else

Requirements

• Core: Python 3.10+, SQLite 3.35+ (FTS5)
• Semantic search (optional): pip install sqlite-vec + one of: GOOGLE_API_KEY / OPENAI_API_KEY / pip install sentence-transformers

Files

File	Lines	Description
__init__.py	630	MemoryProvider implementation
store.py	652	SQLite/FTS5 backend
condenser.py	399	Fact condensation engine
embedding_providers.py	381	Gemini/OpenAI/Local providers
embeddings.py	367	Provider-agnostic vec search
plugin.yaml	7	Plugin metadata
README.md	257	Full documentation
SKILL.md	84	Agent-facing skill docs

Testing

Tested locally with:

• Store CRUD + FTS5 search ✅
• Condensation pipeline (group/dedup/prioritize) ✅
• Semantic search with Gemini embeddings ✅
• Graceful degradation when provider unavailable ✅
• Thread safety (3 concurrent writers) ✅
• Full MemoryProvider integration (prefetch, hooks, tool calls) ✅

[alt-glitch]

This must be a standalone plugin repo now, see CONTRIBUTING.md

[labaznov]

ohh sry just a moment, it seems i will did it again, cancel plz

[labaznov]

Done! Moved to a standalone plugin repo as per CONTRIBUTING.md:

👉 https://github.com/labaznov/hermes-enhanced-memory

Includes:

• Two-tier fact storage with auto-condensation
• FTS5 + semantic vector search (sqlite-vec)
• Pluggable embedding providers (Gemini/OpenAI/Local)
• 187 tests, 78% coverage
• Diplodoc documentation (7 pages)
• pyproject.toml for pip install

Install: cd ~/.hermes/plugins && git clone https://github.com/labaznov/hermes-enhanced-memory.git enhanced_memory