How Hermes Agent's Memory Actually Works: MEMORY.md, Sessions, and Honcho

Most AI agents forget everything between sessions. Hermes has three tiers of memory — and it manages them autonomously.

Tier 1: MEMORY.md — Always Present

Two files injected into every session's system prompt:

• ●MEMORY.md (2,200 chars) — Agent's personal notes: environment facts, project conventions, lessons learned
• ●USER.md (1,375 chars) — Your profile: name, preferences, communication style, workflow habits

The agent adds, replaces, and removes entries on its own. When memory is full, it consolidates entries to make room.

══════════════════════════════════════════════
MEMORY [74% — 1,628/2,200 chars]
══════════════════════════════════════════════
User's project: ~/code/webapp (Python/FastAPI + React)§
Ubuntu 24.04, Docker, uv, nix§
staging: 10.0.1.50:2222, key ~/.ssh/staging_ed25519

Tier 2: Session Search — On Demand

Every conversation (CLI, Telegram, Discord) is stored in SQLite with FTS5 full-text search. The agent can search past conversations weeks later:

session_search(query="Keycloak auth migration")
→ Found 3 sessions from March 18-21:
  "...evaluated Keycloak vs Auth0 vs Supabase..."

Tier 3: Honcho — Deep User Modeling

7 external memory providers including Honcho's dialectic user modeling, OpenViking knowledge graphs, Mem0, and more. These add semantic search, automatic fact extraction, and cross-session understanding.

See It In Action

Watch Hermes recall a conversation from 2 weeks ago, update its memory, and evolve its understanding of the user:

Try Hermes with its full memory system: