Skip to main content
← All use cases
memory_basics.pymulti_agent_shared_space.py

​
The scenario

Your coding agent is excellent for an hour and then the session ends. Next time it re-proposes the pattern you rejected, re-introduces the bug you fixed in March, and re-asks why the module is structured that way. The codebase is in its context. The reasoning about the codebase is not — that lived in conversations, pull request threads and a decision nobody wrote down.

​
Step 1 — Connect over MCP

The fastest path is no code at all. Point the agent at our MCP endpoint and it gets record, retrieve, reason and list_spaces as tools.
MCP tools load at session start, so restart the agent after adding it. Full setup, including keyless OAuth, is on the MCP page.

​
Step 2 — Give it the habit, not just the ability

MCP gives an agent the ability to remember. It will not reliably do so unless told when. Our Agent Skill installs that habit — recall before a task, record after it — without you writing prompt scaffolding.

​
Step 3 — Record decisions, not diffs

Git already has your diffs. What it does not have is why.
Scoping by agent_id keeps one agent’s working notes separable while leaving the decisions readable by everyone — the same shape as Two agents, one space.

​
Evals

  1. Record three decisions. Start a fresh session. Ask the agent why one of them was made, and check it retrieves rather than reasons from the code.
  2. Ask about something never recorded. It should say so — an agent that confabulates a decision is worse than one with no memory.
  3. Re-run a task the agent previously got wrong. The correction should surface.
  4. Check the write actually happened: list_memories after the session, rather than trusting that the tool was called.

​
Guardrails

Never record secrets. Agents paste freely, and a key or connection string written into memory is retrievable by everyone with access to the space and will resurface in later prompts. Keep credentials out of the content you store.
  • A shared team space needs a convention for what is worth recording. Without one it fills with restatements of the code, which crowds out the decisions.
  • Record the rejected option and why. “We use X” is weak; “we rejected Y because Z” stops the next agent proposing Y.
  • MCP tools only load at session start — adding the server mid-session exposes nothing until you restart.

​
Built from

Both are complete, runnable scripts: memory_basics.py, multi_agent_shared_space.py.