Spark Memory

Intelligence layer that compounds. Records decisions, reflects overnight, detects patterns, captures sessions automatically — your agent gets smarter every day.

Zellin@zellin

Install

openclaw plugins install clawhub:@zellin/spark-memory

🧠 Spark Memory — Persistent Intelligence for OpenClaw Agents

Your agent forgets everything when the session ends. Spark fixes that.

Spark gives your OpenClaw agent a memory that compounds. Not flat storage — a brain that reflects, learns patterns, and gets smarter every day.

What Makes Spark Different

Feature	memory-core	memory-lancedb	Spark
Store memories	✅ (files)	✅ (vectors)	✅ (vectors)
Semantic search	❌	✅	✅
Importance scoring	❌	Static 0.7	Dynamic (1-10)
Nightly reflection	❌	❌	✅ Synthesizes patterns
Recency × Relevance scoring	❌	❌	✅ Stanford formula
Cross-session learning	❌	✅	✅ + reflection
Episode types	❌	Basic categories	7 types
Multi-user support	❌	❌	✅ Org-scoped
Survives device loss	❌	❌	✅ Cloud-backed
Selective capture	❌	Trigger-based	✅ Intelligence-based

Install

bash

openclaw plugins install @zellin/spark-memory

Then configure in your OpenClaw settings:

json5

{
  plugins: {
    slots: { memory: "spark-memory" },
    entries: {
      "spark-memory": {
        config: {
          apiKey: "your-spark-api-key",
          orgId: "your-org-id"
        }
      }
    }
  }
}

Setup

Set these environment variables (or add to your OpenClaw config):

bash

export SPARK_API_KEY="your-api-key"    # Get one at https://zellin.ai
export SPARK_ORG_ID="your-org-id"       # Created when you sign up

How It Works

Session starts → Spark recalls relevant memories from previous sessions
During conversation → Spark records important facts, preferences, decisions
Session ends → Spark saves a conversation summary
Every night → Spark reflects on accumulated memories, synthesizes patterns, detects recurring behaviors
Next session → Your agent is smarter than yesterday

Day 1: knows nothing. Day 30: knows your preferences, your clients, your patterns. Day 365: knows more about your operations than you remember yourself.

Architecture

Based on Stanford's "Generative Agents" paper with extensions:

Memory Stream (episodes) → Reflection (synthesis) → Patterns (recurring behaviors)
Retrieval scored by Recency × Importance × Relevance
Org-scoped: multiple users feed one shared memory
Cloud-backed: memories survive device changes

Free Tier

100 episodes/month
Basic recall
Nightly reflection included
Sign up at zellin.ai

License

MIT