Multi-Agent Memory Management: Shared State Without Chaos
NEO added a memory layer for agent teams: short-lived context, longer-lived facts, and clean handoffs so swarms stay readable and do not stomp each other’s state.
Problem Statement
We asked NEO to expose structured read/write APIs for agent memory, with namespaces, TTLs, and conflict rules. Multi-agent apps should not corrupt shared context or leak secrets across sessions.
Solution Overview
- Scoped stores: Per-agent, per-session, and shared pools.
- Policies: Clear rules for who reads and writes which keys, plus hooks for redacting PII.
- Compaction: Summarize older turns so you stay inside context limits.
Workflow / Pipeline
| Step | Description |
|---|---|
| 1. Session start | Allocate memory scopes and retention policy |
| 2. Agent turn | Read relevant slices; append new facts and tool outputs |
| 3. Handoff | Transfer summaries or pointers between agents |
| 4. End | Persist long-term memory; purge ephemeral keys |
Repository & Artifacts
Generated Artifacts:
- Memory service with pluggable backends (file, Redis, vector DB)
- Policy hooks for compliance-sensitive deployments