Introduction

HEBBS is a cognitive memory engine for AI agents. It gives agents temporal reasoning, causal chains, analogical transfer, consolidation, and decay as a single primitive. Deploy on your infrastructure, connect via CLI or SDK, and your agents get memory that understands context, catches contradictions, and learns patterns over time.

The Real Problem

Before HEBBS, “agent memory” meant similarity search. Every solution shares the same architecture: embed a question, query a vector database, return the nearest chunks, hope the answer is in there. This has hard limits:

• One retrieval mode. Ask “what happened with Acme?” and similarity search returns chunks that mention Acme. It doesn’t walk a timeline. It doesn’t trace cause and effect. It doesn’t transfer patterns from another domain.
• No native cognition. Decay, reinforcement, consolidation don’t exist in vector databases. Old noise accumulates alongside new signal forever. Agents don’t learn from their memories; they retrieve from them.
• Stitched together. A production memory stack typically means a vector database, a key-value store, a graph database, an LLM API for extraction, and glue code. Each is a failure mode.

How HEBBS Is Different

• Four recall strategies: similarity, temporal, causal, and analogical. Each backed by a purpose-built index. One call, all strategies available.
• Native consolidation: the reflection pipeline clusters related memories, proposes insights via LLM, validates them, and stores consolidated knowledge with lineage back to sources.
• Automatic decay and reinforcement: memories lose importance over time; frequently recalled memories get reinforced through Hebbian learning.
• Contradiction detection: when new memories conflict with existing ones, HEBBS flags them automatically.
• Configurable scoring weights: every recall blends relevance, recency, importance, and reinforcement. One parameter change shifts behavior.
• Measurable recall quality: HEBBS with the right strategy is 37% more precise than similarity-only search. Tuned weights improve precision by 50-70% over defaults. We measure this with domain-specific evals, not vibes.
• Drop-in agent skill: install the HEBBS skill for Claude Code or OpenClaw. No SDK integration or glue code required.

How It Works

1. Deploy the server on your infrastructure (Docker, one line)
2. Create workspaces for your teams or use cases
3. Upload files via CLI, SDK, or GitHub Actions auto-sync
4. Your agents recall using the right strategy for each question

Files are automatically split into propositions, embedded, and indexed with entity relationships. The entities/ folder convention auto-scopes content to the right entity. Unchanged files are skipped on re-upload.

Architecture at a Glance

Layer	What it does
REST + gRPC API	9 operations: remember, recall, revise, forget, prime, subscribe, reflect, insights, reflect_policy
Embedding	OpenAI, Anthropic, or local ONNX models
HNSW Index	Sub-10ms similarity search at millions of memories
B-tree Index	Temporal range scans by entity + timestamp
Graph Index	Causal edge traversal with bounded depth
Contradiction Detection	Automatic conflict surfacing between memories
Reflection Pipeline	Background LLM-powered insight consolidation
Decay Sweep	Automatic pruning, bounded memory growth

Who Is HEBBS For

• Enterprise teams building AI agents for CRM, support, legal, finance, or any domain where agents need to remember context across interactions.
• AI agent builders who need memory beyond similarity search: temporal reasoning, causal chains, analogical transfer, and consolidation.
• Teams that want zero-config memory: install the agent skill and the agent gets cognitive memory without writing infrastructure code.
• On-prem deployments where data cannot leave the network.

What’s Next

Head to the Quickstart to deploy the server and store your first memory in under 5 minutes.

Or read Key Concepts to understand the memory model, recall strategies, and reflection pipeline before diving in.