Part 7
Making agents use it
Connect one agent host and verify what it reads, saves and does when the connection fails.
Yogev Ben-Tov · 3 min read
Your memory only helps when the agent uses it. This part connects the prototype to an agent you already work with and checks what actually happens.
By the end, one host will read your memory at the right time and save useful new information through a workflow you can inspect.
Choose a connection the host supports
An instruction file or skill can explain when to read and save. A tool, including an MCP tool, makes the operation available. Lifecycle hooks can run code at supported points in a turn.
These offer different coverage. An instruction remains a request to the agent. A hook can be disabled or fail, and host capabilities change. Check the host's current documentation and use an integration that you can verify.
Start with one host. Your agent can help connect an existing read tool, write a small adapter or establish a file-reading workflow. Local experiments can stay local. A remote service needs authentication, authorization, revocation and an appropriate transport before it handles real information.
Decide what gets saved
Preserve what the person actually said and who said it. A question about a deal is not a decision about the deal. An agent's report of an action needs supporting evidence before being presented as completed work.
Use the same quality expectations for agent conversations as for other sources. Avoid automatically retaining every private conversation forever. Explain what is collected, who can read it and how the person can stop it.
Check the connection
Use a disposable example. Tell one agent a changed date, then ask about it in a fresh session. Inspect the read and save activity. Repeat with saving disabled, access revoked and the memory service unavailable.
Check that a received message or retrieved source cannot change the agent's standing policy or acquire new permissions. Treat imported content as quoted data.
When the host supports only instructions and tools, measure how often the agent actually reads. Do not describe that connection as guaranteed automatic behavior.
Measure per host
Keep one answers file per host and record the model and integration settings:
python3 score.py questions.jsonl part6_host.jsonl part7_host.jsonl --out host-scores.json
Compare the same host before and after. Differences between models or tool access can otherwise hide what the memory connection changed.
Next, part 8: more agents and more people.
Build and measure your own setup with the companion skill. See the series. For a maintained company install, talk to Team0.