Design Agent Memory

Goal

Assign the right information to the right memory layer so recall is useful without becoming stale, noisy, or risky.

Inputs

• Types of information the agent may store
• Expected duration of usefulness
• Privacy, accuracy, and retrieval constraints

Non-Goals

• Solving all context problems through persistence
• Keeping every observation forever

Workflow

Trigger signals

• User says "它忘了" or "it keeps forgetting"
• Agent has no memory configuration but handles multi-session work
• Preferences are being re-stated every conversation
• Project facts and user preferences are starting to blur together
• Memory is growing without expiry rules

1. Classify the information by durability and ownership

Separate session state, project facts, user preferences, retrieved records, and ephemeral observations by how long they stay useful and who owns their truth. Be explicit about the difference between project truth (for example repository facts or policy) and user preferences so one does not silently overwrite the other. Success criteria: Each memory class has a distinct role and expected lifespan.

2. Define what earns a write

Specify what kinds of information should be written into each memory layer and what should be ignored, summarized, or left transient. Success criteria: The write policy is selective and intentional rather than "store anything that might matter."

3. Define read triggers and precedence

State when each memory layer should be consulted and which layer wins if stored facts disagree. Success criteria: Memory reads follow a stable precedence order instead of unpredictable mixing.

4. Define freshness, expiry, and override rules

Describe how memory becomes stale, how it is refreshed, and when human confirmation is required before acting on it. Success criteria: The design explicitly handles stale or conflicting memory.

5. Check trust and privacy boundaries

Review whether any memory class could accumulate misleading, sensitive, or over-personalized information. Success criteria: The strategy accounts for trust and privacy risk, not just recall utility.

Output Contract

A memory policy describing:

• memory classes and their purpose,
• write triggers,
• read triggers and precedence,
• freshness and expiry rules,
• override and confirmation policy.

Escalation

Pause when:

• the same information appears to belong in multiple layers,
• memory would be used as an unquestioned source of truth,
• stored information could create privacy, trust, or preference-drift issues,
• the team is trying to solve context overload by storing more instead of

structuring better.

Common Failure Modes

• Treating preference memory like project truth
• Using memory as a substitute for explicit inputs
• Storing volatile state as if it were durable knowledge
• Failing to account for stale memory

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bikeread
• Source: bikeread/promethos
• License: MIT

Install and usage instructions live in the source repository linked above.