Hebbs is a single Rust binary with embedded storage. No Postgres. No Redis. No Pinecone. No infra to manage. The core operations:
- remember — store experiences with automatic importance scoring.
- recall — four retrieval strategies (similarity, temporal, causal, analogical), all sub-10ms p99.
- reflect — consolidate raw memories into distilled insights with full lineage tracking. This is how agents go from "I remember what you said" to "I know what I've learned"
- revise — update beliefs without appending contradictions.
- forget — GDPR-safe targeted erasure, plus active pruning of stale knowledge.
Everything is 100% configurable ( embedding model, half life decay period, ef_search, analogical_alpha ), relevance, importance, recency and reinforcement weights, but ships with smart defaults that worked well for me building what I was focused on - so you don't need to tune anything on day one.
Built for both edge ( robots ) and server. Run it embedded in your agent process for zero-latency local memory, or run the server for multi-agent setups. Reflect syncs only distilled insights to a master server - so your sub-agents stay lightweight while the fleet shares institutional knowledge.
The thesis: agents that learn compound over time. Call 500 is dramatically better than call 1 - not because the agent has more data, but because it has distilled knowledge. No current system delivers that. You need to duct-tape solutions and that path is painful and tweak to make it yours as you live on their default configs.
[+] [-] sch00lb0y|16 days ago|reply
I'm glad to see someone paving the way for the future.
It would have been great if it was like CMD line tool instead of server. Since agents are good at interacting with cli tools
[+] [-] paragarora|19 days ago|reply
- remember — store experiences with automatic importance scoring.
- recall — four retrieval strategies (similarity, temporal, causal, analogical), all sub-10ms p99.
- reflect — consolidate raw memories into distilled insights with full lineage tracking. This is how agents go from "I remember what you said" to "I know what I've learned"
- revise — update beliefs without appending contradictions.
- forget — GDPR-safe targeted erasure, plus active pruning of stale knowledge.
Everything is 100% configurable ( embedding model, half life decay period, ef_search, analogical_alpha ), relevance, importance, recency and reinforcement weights, but ships with smart defaults that worked well for me building what I was focused on - so you don't need to tune anything on day one.
Built for both edge ( robots ) and server. Run it embedded in your agent process for zero-latency local memory, or run the server for multi-agent setups. Reflect syncs only distilled insights to a master server - so your sub-agents stay lightweight while the fleet shares institutional knowledge.
The thesis: agents that learn compound over time. Call 500 is dramatically better than call 1 - not because the agent has more data, but because it has distilled knowledge. No current system delivers that. You need to duct-tape solutions and that path is painful and tweak to make it yours as you live on their default configs.