[MEM]...[/MEM] block. When the agent needs a value, the fine-tuned Hermes3 reader model performs implicit frequency decomposition — no dictionary lookup, no semantic search, no vector similarity.
The one-line positioning
Honcho remembers what kind of person you are. Spectral Memory remembers your token budget is 1024.
What you can store
User facts
Name, GPU, OS, working directory, affiliation — anything that should persist across sessions.
Preferences
Response style, verbosity, code style, output format — exact values, not summaries.
Task state
Current task, next task, blockers, deadlines, priorities — survives compression intact.
Project metadata
Model target, token budget, last result, accuracy — precise numbers, not narratives.
How it compares
| Spectral Memory | Plaintext (MEMORY.md) | Narrative (Mem0) | Vector (RetainDB) | |
|---|---|---|---|---|
| Exact recall | ✅ | ❌ degraded by compression | ❌ inferred | ⚠️ approximate |
| Compression-safe | ✅ via /flush | ❌ | ❌ | ❌ |
| Prompt injection resistant | ✅ | ❌ | ❌ | ❌ |
| Private by construction | ✅ | ❌ | ❌ | ❌ |
| Infrastructure required | API only | None | Embedding model | Vector DB |
Base URL
Quickstart
Encode your first fact in under 5 minutes.
Hermes Plugin
Install the agent plugin and wire up automatic memory injection.