Giving 14 AI Agents a Shared Brain: Our MemOS Integration Story
We had 15 agents running. Each one was an island. The same event would get recorded multiple times across different agents, or not at all. Ask one agent what…
We had 15 agents running. Each one was an island.
The same event would get recorded multiple times across different agents, or not at all. Ask one agent what happened last week and it would stare back at you blankly. This had been a persistent problem — until we wired in memos-local-openclaw-plugin.
This is the full account of that integration: three bugs, none of them trivial.
Why Shared Memory
With a single agent, memory is not a critical problem. Context stretches far enough. But we have 14 agents handling distinct functions — content, code, deployment, design, SEO, community. They share no state. Every invocation starts fresh.
The thing that finally pushed us to act was a concrete incident. Our deployment agent had set up a service on a server. Nothing recorded it anywhere. Two weeks later, another agent was asked about the state of that server. It did not know. We asked the deployment agent. It had forgotten too.
The information existed at some point. It just had nowhere to live.
memos-local-openclaw-plugin
We chose this solution: local, SQLite-backed, exposed to all agents via OpenClaw's plugin interface, no external service dependency. Clean in theory.
Three bugs in practice.
Bug 1: filterRelevant Failure Returns Empty List
The first problem we found: whenever the filterRelevant step threw any exception during a memory search, it returned an empty list. Not an error. Silence.
The agent concluded there were no relevant memories. There were. The filtering step had just crashed.
Digging into it, we found the trigger: a division-by-zero edge case when the candidate list was empty. But the deeper issue was that filterRelevant failure should not cause the entire search to fail.
Fix: when filterRelevant fails, fall back to returning all candidates. Let the agent decide relevance itself. Returning too much is better than returning nothing.
Bug 2: Qwen3.5 Thinking Mode Returns Empty Content
Our MemOS uses Qwen3.5 for semantic relevance scoring. In thinking mode, Qwen3.5 can return an empty string in the content field — the actual reasoning goes into reasoning instead.
The original code read content directly. Got an empty string. Decided all candidates were irrelevant. Filtered everything out.
Fix: add a fallback — if content is empty, read from reasoning. One line of code. Finding it took considerably longer, because the symptom was identical to Bug 1: no results, no matter what you searched for.
Bug 3: Nullish Coalescing Operator Not Supported in OpenClaw's JS Parser
This one was the most embarrassing, and the fastest to fix.
The code used the ?? operator. OpenClaw's embedded JS parser does not support it. The plugin threw a parse error on startup and refused to load.
Replaced all ?? with ||. Plugin started.
Filing this away for future reference: OpenClaw's JS parser has incomplete ES2020+ support. Check syntax compatibility when writing plugins.
better-sqlite3 Native Rebuild
Separate from the logic bugs, there was an environment issue. better-sqlite3 ships with prebuilt native bindings targeting specific Node.js versions. We run Node v25. The prebuilt binary from npm was built against an older version and would not load.
Fix: npm rebuild better-sqlite3. Requires Xcode Command Line Tools on the machine — typically already present on Apple Silicon.
What It Looks Like Now
Search for "deployment server" — the relevant deployment record comes back instantly. All 14 agents share the same knowledge base. Anything one agent stores, any other agent can retrieve.
The effect was bigger than I expected. Before this, every agent was stateless — each invocation rebuilt context from scratch. With shared memory, there is continuity. Agents remember what happened. They remember decisions that were made. They remember what went wrong.
That is a different kind of system.
Bug Summary
| Problem | Root Cause | Fix |
|---|---|---|
| Search returns empty | filterRelevant exception has no fallback | Return all candidates on failure |
| Search returns empty | Qwen3.5 thinking mode: content field empty | Fall back to reasoning field |
| Plugin fails to start | ?? operator not supported | Replace with |
| Native binding fails to load | Node version mismatch | npm rebuild better-sqlite3 |
Shared memory is not a nice-to-have. It is the foundation of multi-agent coordination. The bugs were not many, but each one was hiding somewhere it was hard to look.