Refine Memory and Tool Usage

After instructions and workflow, the next two levers worth pulling are the memory layer and the tool catalog. Both are easy to expand thoughtlessly and both quietly degrade agent quality when they are. The refinement discipline is to add little, scope tight, and let evaluator signals tell you when to prune.

Refining the tool catalog

Microsoft Foundry's tool catalog includes built-ins (web search, file search, memory, code interpreter), custom functions, and MCP servers; a Toolbox lets you curate a single MCP-compatible endpoint of approved tools per agent. The signals that drive refinement:

| Signal | Refinement | | --- | --- | | Tool-call accuracy drops after adding tools | Scope: hide tools the agent doesn't need, or split into a smaller per-agent Toolbox | | Agent often picks the wrong tool from two similar ones | Improve tool descriptions and parameter names; consider merging | | Tool returns errors the agent ignores | Document expected errors in the tool description; add handling in instructions | | Tool succeeds with wrong args | Tighten the schema; add input validation in the wrapper |

Rule of thumb: every tool the agent never uses on a real task is a tool that lowers selection accuracy on the ones it does use.

Refining the memory layer

Memory turns into a context-failure machine when nobody owns it. Memory hygiene rules:

• Write less: only persist facts that have to outlive the conversation.
• Update on write: if a fact has a canonical owner (account record, profile), refresh from source on read instead of caching forever.
• Expire on time: every memory entry has a TTL, even if the TTL is "until the next session".
• Never persist secrets: tokens, credentials, and PII that the agent saw in passing should not live in durable memory.
• Prefer authoritative source: when memory disagrees with a system of record, the system of record wins.

When a failure is "agent used a stale fact", it is almost always a memory hygiene problem, not a model problem.

Scope is a quality lever, not only a security lever

Granting a tool the full surface ("the whole org", "any file path") inflates the agent's choice space and the blast radius of mistakes. Scoping down to the minimum surface needed for the task improves tool-call accuracy and limits damage when something goes wrong — which is why Foundry exposes scope as a first-class configuration on connected tools and MCP servers.

Quick check

Where this shows up on the exam

Two recurring shapes: (1) tool-call accuracy drops after a catalog expansion — the answer is to scope down or curate a Toolbox; and (2) the agent answers using a stale value — the answer is memory hygiene (expiry, write-through, source-of-truth precedence), not a model swap.