Seed Memory Design Principles

What to Include and What Causes Harm

Seed memory is background information actively defined by the user that is automatically injected into the system at the start of every conversation. Designed well, it allows the AI to understand from the first sentence who the user is, what they are working on, and how they prefer to communicate. Designed poorly, it constrains the AI's flexibility and can introduce interference in unrelated contexts.

Seed memory design and maintenance is a core component of AI long-term memory system architecture. If seed memory entries drift or become inaccurate over time, see AI Memory Corruption: Identification and Correction.

Information Types Worth Including in Seed Memory

Stable background facts: Occupation, industry, working language, institution type. These change infrequently and including them avoids repeating the same background in every conversation.

Communication style preferences: Response length (brief or detailed), format (bullet points or prose), language requirements (formal written Chinese or Traditional Chinese), expressions to avoid. AI style calibration is highly effective, and these preferences are typically persistent.

Standing work context: Ongoing long-term projects, commonly used tools or platforms, regularly handled document types. For example: "Currently responsible for back-end development of System X, primarily using Next.js and Neon Postgres." This type of information allows the AI to enter a working state directly for relevant tasks, without requiring environment explanation.

Specific constraints or limitations: "Legal questions must explicitly note that I should consult a lawyer." "Financial advice is limited to general principles, no specific operational recommendations." These constraints help with boundary management, especially in high-stakes decision contexts.

Information That Should Not Be in Seed Memory

Temporary states: "My workload is extremely heavy this week, please simplify responses as much as possible" is a temporary situation. Once placed in seed memory, it remains active weeks later and affects output quality in contexts unrelated to workload. Temporary requirements should be stated in the current conversation, not written to persistent memory.

Contradictory instructions: "Responses should be brief" and "Responses should be comprehensive and detailed" existing simultaneously produces unpredictable model trade-offs, and different tasks may require entirely different levels of depth. These preferences should be specified per task type within the conversation, not set as global rules.

Too much low-leverage detail: Seed memory information shares the same context space as actual task content. Seed memory exceeding 2,000 words dilutes the model's attention on the current query, causing background information to crowd out task-relevant weight. The principle is to include only high-leverage facts that directly affect most conversations, not a comprehensive personal encyclopaedia.

Update Cadence

Seed memory should be treated as a periodically maintained document, not a system parameter set once and never revised. Update immediately when professional role changes; actively review every 3 to 6 months to delete or adjust entries that are no longer applicable; if an entry has not played a role in conversations over the past few months, reassess whether it is necessary.

Summary

The design principle for seed memory is: stable, high-leverage, and non-contradictory. Information included should be background facts that are effective in the vast majority of conversations, not every personal detail that might be relevant. Well-trimmed seed memory consistently outperforms content-rich but lengthy versions in practice.

Further reading: Five Practical Business Uses of Chat History