Developer Tom Dörr released a three-layer hybrid memory system for multi-agent OpenClaw setups on GitHub under the MIT license on June 30, 2026, combining files, a graph, and shared storage.
June 30, 2026 · MIT License
A Three-Layer Hybrid Memory System for Multi-Agent OpenClaw
"openclaw-graphiti-memory" combines private files, shared files, and a temporal knowledge graph — balancing the transparency of human-readable memory with graph-based cross-cutting search to fight forgetting and context contamination.
3
memory layers: files, shared files, graph
<$1
per month at 20-agent scale (gpt-4.1)
$0
search cost — local Neo4j queries
The Three Layers, Stacked
Layer 1 · Private Files
QMD — vector + BM25
Each agent's own memory/ directory, queried with hybrid full-text and vector search.
↓
Layer 2 · Shared Files
read-only, orchestrator-managed
A shared/ directory — user profiles, agent roster, infrastructure info — mounted to each agent via symlinks.
↓
Layer 3 · Knowledge Graph
Graphiti + Neo4j
A shared graph strong on temporal facts; each agent writes to its own group but can search across groups.
Why isolate per-agent groups?
Separating workspace from memory is the key to clean multi-agent operation — each agent lives in a dedicated group to prevent cross-mixing of information.
Forgetting
No more per-session zero-starts
Contamination
Per-group isolation across agents
Cross-cutting
Temporal & invoice search across groups
Strengths
Human-readable, editable file memory
Temporal tracking of config changes
Near-zero running cost; free local search
Limitations
Cost of Graphiti entity extraction
Overhead of symlink management
Cross-group accuracy depends on the model
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…