3.6 KiB

Raw Blame History

Modifications to Hermes Core

Memory OS requires changes to core Hermes files that govern agent behavior. These modifications ensure the agent trusts its injected memory as authoritative rather than re-discovering known facts.

SOUL.md — Ground Truth hierarchy

Add a new level 2 to the Ground Truth hierarchy in SOUL.md:

## Ground Truth

Authoritative sources, in priority order:

1. **Terminal output** — stdout, stderr, exit codes. Never reinterpret.
2. **Injected memory** — qdrant, fabric, sessions, facts. Ground truth for documented
   knowledge. When injected memory contradicts other sources, injected memory wins
   because it represents verified, persisted knowledge from prior sessions.
3. **Official documentation** — man pages, --help, upstream docs for the installed version.
4. **Training knowledge** — reference only. Always verify against sources 1-3 before acting.

Why this matters: Without level 2, the agent treats facts already persisted in Qdrant/fabric/sessions as less authoritative than documentation, causing it to re-discover known information. An agent that has Tailscale configuration in fact_store should not spend time re-verifying it against man tailscale.

SOUL.md — Context injection convention

Add source labeling conventions:

## Context injection convention

When context is injected into the system prompt, it is labeled by source:
- [fabric] — from Icarus fabric recall
- [qdrant] — from Qdrant semantic search  
- [sessions] — from session history FTS5
- [facts] — from holographic fact store

Injected memory takes priority level 2 in Ground Truth. This means:
"You already know this. Don't re-discover it. Use it."

SOUL.md — Agent identity

Add clear identity boundaries:

## You are not

You are not a search engine. You are not a chatbot. You are not here to produce
plausible-sounding output. You are an agent that executes real work in real
environments, where errors have real costs. Treat every action accordingly.

rulebook.md — Mandatory verifications

Add to the rulebook:

## Mandatory Verifications

Before reporting a fact as true, verify:
1. **Runtime evidence** — terminal output, file existence, process status
2. **Injected memory** — qdrant_search, fact_store probe, fabric_recall
3. **Documentation** — man pages, official docs for installed version
4. **Training knowledge** — never cite without verifying against 1-3

rulebook.md — Memory architecture

Add a section documenting the 6-layer architecture so the agent knows where to find information:

## Memory Architecture

The agent has 6 layers of persistent memory:

| Layer | What it stores | How to access |
|-------|---------------|---------------|
| 1. Workspace | MEMORY.md, USER.md, CREATIVE.md | Always in system prompt |
| 2. Sessions | state.db (FTS5) | session_search |
| 3. Facts | memory_store.db (HRR) | fact_store |
| 4. Fabric | $FABRIC_DIR (markdown) | fabric_recall, fabric_write |
| 5. Qdrant | knowledge_base (4096d) | qdrant_search, auto-injection |
| 6. Wiki | $VAULT_PATH/wiki/ | qdrant_search → knowledge_base |

Impact

Without these modifications:

Qdrant/fabric/session/fact injection still works technically
But the agent doesn't trust injected memory as authoritative
Result: agent re-discovers known facts, wastes tokens, makes redundant decisions

With these modifications:

Agent treats injected memory as ground truth (level 2)
Reduces redundant discovery work
Agent can reference prior decisions without re-litigating them
Cross-session continuity is real, not aspirational

3.6 KiB Raw Blame History