mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-25 00:51:20 +00:00
Brain-analogous AI memory with associative networks, ACT-R/FSRS decay, spreading activation, and LLM-powered dream consolidation. Features: - 5 tools: cc_remember, cc_recall, cc_todo, cc_message, cc_health - Automatic prefetch (relevant memories injected before each turn) - Background sync (significant user messages auto-stored) - Session end summaries - Built-in memory mirroring (MEMORY.md writes → CerebroCortex) - Delegation capture (subagent results stored as memories) - Multi-agent messaging (send_message/check_inbox) - Zero cloud dependencies — SQLite + ChromaDB + igraph, all local Install: pip install cerebro-cortex Repo: https://github.com/buckster123/CerebroCortex Closes #7896 |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| plugin.yaml | ||
| README.md | ||
CerebroCortex Memory Provider
Brain-analogous AI memory for Hermes Agent with associative networks, biologically-inspired decay, and dream consolidation.
What it does
CerebroCortex replaces simple key-value memory with a system modeled on how biological memory actually works:
- 6 memory types — episodic, semantic, procedural, affective, prospective (TODOs), schematic (patterns)
- Associative graph — memories link together via 9 typed/weighted link types. Search spreads activation through the graph, surfacing related memories you didn't explicitly search for
- Natural decay — ACT-R power-law activation + FSRS spaced-repetition. Frequently-used memories stay sharp, noise fades automatically
- Dream engine — Offline LLM-powered consolidation: pattern extraction, schema formation, pruning, cross-domain discovery
- Multi-agent messaging — send_message/check_inbox for agent-to-agent communication through shared memory
All data is local (SQLite + ChromaDB + igraph). No cloud services required. Runs on hardware as minimal as a Raspberry Pi 5 (4GB).
Install
pip install cerebro-cortex
Setup
hermes memory setup
# Select "cerebrocortex" from the list
Or manually in ~/.hermes/config.yaml:
memory:
provider: cerebrocortex
Tools
| Tool | Description |
|---|---|
cc_remember |
Store a memory (auto-classifies, deduplicates, links) |
cc_recall |
Semantic search with graph activation and decay scoring |
cc_todo |
Store/list/resolve TODOs and reminders |
cc_message |
Cross-agent messaging (send/inbox) |
cc_health |
System health and statistics |
Automatic features
- Prefetch — Before each turn, relevant memories are recalled and injected as context
- Sync — Significant user messages are auto-stored in background
- Session end — Session summary saved automatically
- Memory mirror — Built-in MEMORY.md/USER.md writes are mirrored to CerebroCortex for searchability
- Delegation capture — Subagent task/result pairs stored as memories
Configuration
| Key | Description | Default |
|---|---|---|
agent_id |
Agent identifier for multi-agent setups | HERMES |
data_dir |
Data directory path | ~/.cerebro-cortex/ |
Set via hermes memory setup or in config.yaml:
plugins:
cerebrocortex:
agent_id: HERMES
data_dir: ~/.cerebro-cortex/
Multi-agent
Multiple Hermes instances can share a CerebroCortex store:
# Instance 1
plugins:
cerebrocortex:
agent_id: HERMES-PRIMARY
# Instance 2
plugins:
cerebrocortex:
agent_id: HERMES-RESEARCH
Agents communicate via cc_message:
cc_message(action="send", to="HERMES-RESEARCH", content="...")cc_message(action="inbox")