Getting Started
Install HEBBS, start the server, and store your first memory in under 5 minutes.
HEBBS Documentation
The cognitive memory engine for AI agents. File-first local vaults. Four recall strategies. Native consolidation. Drop-in agent skill.
Vault (Local Mode)
File-first cognitive memory. Index your markdown files, watch for changes, and recall locally with no server required.
Python SDK
Async Python client with a clean, Pythonic API. No protobuf in the public surface.
TypeScript SDK
Async Node.js gRPC client with full type safety. npm install @hebbs/sdk.
API Reference
9 operations across 3 groups: Write, Read, and Consolidation. gRPC and REST.
Agent Skill
Drop-in SKILL.md for any compatible agent. No SDK integration required.
Cookbooks
Production-ready tutorials: sales agents, support bots, GDPR compliance, real-time memory surfacing.
Why HEBBS?
Section titled “Why HEBBS?”Before HEBBS, agent memory was similarity search - embed a question, find the nearest vectors, hope for the best. There were vector databases, memory wrappers, and knowledge graphs, but nothing that gave agents temporal reasoning, causal chains, analogical transfer, consolidation, and decay as a single consumable primitive.
HEBBS is the cognitive memory engine that didn’t exist until now:
- Four recall strategies - similarity, temporal, causal, and analogical - each backed by a purpose-built index. 91% precision on temporal queries where similarity-only gets 23%. 78% on causal where similarity-only gets 15%.
- Native consolidation - the reflection pipeline clusters raw episodes, proposes insights, validates them, and stores consolidated knowledge with full lineage back to source memories. Your agent learns, not just retrieves.
- Automatic decay and reinforcement - memories fade over time unless reinforced by access. Signal-to-noise improves naturally. No manual pruning.
- Revision with lineage -
revise()replaces beliefs while preserving the predecessor chain. No contradictory facts coexisting. - Drop-in agent skill - install the SKILL.md file and any compatible agent gets cognitive memory without SDK integration or glue code.
- File-first vault mode -
hebbs initin any directory and your markdown files become searchable memories. A background daemon shares the embedding model across all your vaults. - Contradiction detection - HEBBS automatically surfaces conflicting beliefs in your knowledge base and writes human-readable reports.
- Memory Palace - a built-in web UI (
hebbs panel) for visualizing your memory graph with clusters, edges, and live updates. - One binary, zero external dependencies - and it does all of this in under 10ms at 10 million memories.
Core Operations
Section titled “Core Operations”| Group | Operations | Purpose |
|---|---|---|
| Write | remember, revise, forget | Store, update, and prune memories |
| Read | recall, prime, subscribe | Retrieve context via multiple strategies |
| Consolidate | reflect_policy, reflect, insights | Background learning and knowledge synthesis |