Installing Shared Memory for 14 AI Agents: MemOS Real-World Lessons
The Problem We Were Solving We have 15 agents running. Each one is an island — the same event gets logged multiple times by different agents, or not at all. …
The Problem We Were Solving
We have 15 agents running. Each one is an island — the same event gets logged multiple times by different agents, or not at all. Ask one agent what happened last week and it's blank. The problem had been dragging for a while. We finally addressed it by installing memos-local-openclaw-plugin.
This is the complete bug log from that integration — three bugs, none of them small.
Why Shared Agent Memory Is Hard
When you have one agent, memory is simple: it reads its own files. When you have 15, you need a memory system that handles concurrent reads and writes without corruption, that makes each agent's relevant memory accessible to others, and that doesn't create a single point of failure for the whole system.
Bug 1: Concurrent Write Corruption
Multiple agents writing to shared memory simultaneously produced corrupted entries. The plugin's default configuration assumed single-writer access. Fix: file-level locking around all write operations. Adding the locking increased write latency by about 40ms per operation — acceptable trade-off for data integrity.
Bug 2: Retrieval Scope Confusion
Agents were retrieving memories from the wrong namespace — reading each other's specialized context when they should have been reading shared context only. The plugin's namespace hierarchy wasn't clearly documented. We had to read the source code to understand the expected structure. Fix: explicit namespace configuration in each agent's setup.
Bug 3: Memory Bloat Under Load
After two weeks of production use, the shared memory store had grown large enough that retrieval latency increased significantly. No built-in pruning mechanism. Fix: implemented a nightly cleanup job that archives entries older than 30 days and removes duplicates.
Is It Worth It?
After the fixes: yes. The core use case — agents being able to build on each other's work without human re-briefing — works reliably. Tasks that used to require a human to relay context between agents now hand off cleanly. The three bugs were real work to fix, but the underlying architecture is sound.