Open practice
Agent Memory Protocol
A small, portable routine for coding agents to carry forward what matters without turning an instruction file into a transcript.
What belongs in memory
Store learned context: decisions, pitfalls, working patterns, important project facts, and concise handoffs. Keep stable commands, rules, and conventions in CLAUDE.md, AGENTS.md, or the equivalent instruction file. The two layers are complementary.
01
Recall first
At the start of a task, retrieve memories about the project, goal, and files you are about to change.
02
Store durable lessons
Save decisions, pitfalls, patterns, and handoff context that another session would otherwise have to rediscover.
03
Verify when it helps
When a recalled memory materially improves the work, verify it so future recall can rank it with more confidence.
04
Leave a handoff
Before another agent or session continues, store the outcome and the next useful step.
The portable instruction
Add this intent to the instruction file your coding client reads. Tool names vary by client and provider; use the equivalent recall, store, verify, and handoff tools that are available to the agent.
At the start of each task, recall relevant project memory.
During work, store durable decisions, pitfalls, patterns, and important context.
When recalled memory materially helps, verify it.
Before ending or handing off, store the outcome and the next useful step.Verify a connection
After connecting two agents to the same StremAI project, run this with the first agent and then ask the second agent to recall the same setup check. The second agent should return the stored title and content rather than a fresh guess.
> Use StremAI project memory and test store/recall.
> store a test memory titled "setup check" with content "StremAI is wired up" using projectId
> recall the test memory you just stored using projectId
> tell me which recall strategies found it