Why Agents Forget: Context Windows Grew, but Memory Needs a Fact Store
Context windows grew from tens of thousands to over a million tokens, yet users will not re-explain a project's background from three months ago in every new chat. The common fix is three tiers: raw chat kept verbatim short-term, session summaries holding decisions, and a fact store carrying preferences and project context long-term. Builders get a checklist for stopping agents re-asking what they were already told: memory means distilling conversations into reusable facts, not storing every word. The next race: a standard interface carrying those facts across apps.
At a glance
- 1A bigger context window does not solve cross-session memory
- 2Memory layers separate raw chat from reusable facts
- 3Whoever owns the memory layer owns the agent ecosystem entry point
Why stronger models still feel forgetful
Context windows grew from tens of thousands to over a million tokens, which looks like the problem is solved. In reality users will not re-explain a project's background from three months ago. The window is current vision; memory is historical asset.
How memory layers are structured
Memory is not storing every word — it is distilling conversations into reusable facts.
A common three-tier design:
| Tier | Stores | Retention |
|---|---|---|
| Raw chat | Verbatim | Short |
| Session summary | Decisions | Medium |
| Fact store | Preferences, project context | Long |
A concrete example
When a user says "I dislike long replies" once, that preference should land in the fact store, not be re-inferred every time.
# Pseudo: write preference to fact store
memory.write({ type: "preference", value: "brief_replies" })What to watch next
Whoever turns memory into a standard interface lets agents carry experience across apps. That is the race now.
Related
Full-text Markdown
Copy or download in one click, then paste into notes or feed it to an AI