Building an AI Agent with Persistent Memory
Most AI assistants start fresh every conversation. Hermes Agent doesn't.
The Memory Stack
Hermes uses a three-layer memory system:
1. SQLite FTS5 (Full-Text Search) — Fast keyword search across all past conversations. Every session is indexed and searchable within milliseconds.
2. Vector Embeddings — Semantic similarity search. When you mention a past project, Hermes finds related context even if you don't use the exact same words.
3. Knowledge Graph (gbrain) — Structured relationship storage. Pages, tags, links, and timeline entries form a connected graph of knowledge that persists across sessions.
Implementation
The memory system is built on SQLite with the sqlite-vec extension for vector search:
# Memory is injected automatically at conversation start
memory_context = session_search("relevant past context")
system_prompt += f"\n\nRelevant context:\n{memory_context}"
No external services. No cloud dependencies. Your data stays on your machine.
Why This Matters
- Continuity: Reference past decisions without repeating yourself
- Learning: The agent gets smarter about your preferences over time
- Autonomy: Make complex multi-session plans that actually execute
Try It
pip install hermes-agent
hermes setup
Open source on GitHub. MIT licensed.
Top comments (0)