Knowledge and Memory
Which knowledge should influence an AI system’s next answer?
Start with the first report → See all reports ↓Overview
An AI system can draw on model weights, the current conversation, stored records and live tools. Those sources have different lifetimes and authority. This serial starts by asking where a changing piece of knowledge belongs, then follows two branches: how to decide between conflicting evidence, and how retained information should influence later interactions.
The memory branch moves from what to retain to when that retained information should influence the present interaction. It ends by asking how a correction or deletion can remain effective in later behavior. Each step changes the lifecycle question: choose the observation, justify its reuse, then make a user’s change reach the places where the information still matters.
Stored knowledge leads to two different decisions about use
Place the knowledge
Choose a home and define what a safe change must preserve.
Resolve conflicting evidence
Align claim scope, source independence and unresolved uncertainty.
Manage memory over time
Retain useful context, justify its reuse and make corrections last.
Source adjudication and persistent memory share the placement question, then follow different paths.
Reading order
A recommended path through the enquiry. Each report explains its part; continuation links preserve the original branches.
- Choose where knowledge belongsWhere an LLM System’s Knowledge Lives—and How to Change It Safely
Provides the shared foundation: knowledge placement, scoped change contracts and system-level verification.
Start here - Follow the evidence branchWhen Sources Disagree: A Decision Model for Deep Research
Asks how a research system should combine conflicting material after distinguishing instructions from factual support.
- Begin the memory branchWhat Should an AI System Remember?
Defines what retained observations should preserve so that later interactions can benefit without turning context into a permanent rule.
- Decide when the past belongsWhen Should an AI Use Its Memory?
Focuses the reuse question on the present task, continued validity, appropriate influence and user correction.
Continues What Should an AI System Remember? - Make corrections persistHow to make AI memory corrections and forgetting last
Follows a change beyond its acknowledgment into derived state, work in progress and later ingestion.
Continues When Should an AI Use Its Memory?
Further reading
An earlier architecture-focused examination of accurate memories causing wrong decisions, for a closer look at the failure modes behind reuse.