hawkes.aiKey ideas
- The one idea — Since the model is stateless, agent memory is built outside it, in tiers: short-term (the context window) and long-term (an external store, often a vector DB, the agent writes to and retrieves from across sessions) — memory is engineered, not innate.
- How it is shown — An agent "remembering" you next week — not because the model changed, but because it saved a note to an external store last time and retrieved it now (RAG over its own memory).
- The trap to avoid — Expecting the model to "just remember" across sessions — it's stateless; memory is a system you design (what to save, where, how to retrieve).
- What it sets up — The model has knowledge and memory, but every tool/source is wired in ad hoc — is there a standard plug?
That AI that remembers you from last week? The model itself remembers nothing — it's stateless. Someone engineered that memory: stored notes, fetched back just in time.
The one idea
Since the model is stateless, agent memory is built outside it, in tiers: short-term (the context window) and long-term (an external store, often a vector DB, the agent writes to and retrieves from across sessions) — memory is engineered, not innate.
The model is stateless — it remembers nothing between calls. So when an agent seems to remember you across a session, or even next week, that memory isn't in the model. It's scaffolding you build around it, in tiers. Tier one is short-term, or working memory. It's what's in the context window right now — the current conversation, recent steps, a scratchpad. Fast and immediate, but finite, and gone the moment the window clears. This is where context rot bites. Tier two is long-term memory: an external store the agent writes important things to and retrieves from later — often a vector database, so it's semantic-searchable.
How it works — the demo
An agent "remembering" you next week — not because the model changed, but because it saved a note to an external store last time and retrieved it now (RAG over its own memory).
This one persists across sessions. And retrieving from it is just RAG, run over the agent's own notes. So next week it remembers you not because the model changed, but because it saved a note last time and retrieved it now. Write to the store on the way out; pull the relevant piece back on the way in. Memory is retrieval over a store the agent maintains. People split it further — episodic memory of past interactions, semantic of facts it learned, procedural of how to do things. The names help, but the pattern underneath is always the same: decide what's worth saving, where to put it, and how to fetch it back. The trap: expecting the model to just remember across sessions.
The trap to avoid
Expecting the model to "just remember" across sessions — it's stateless; memory is a system you design (what to save, where, how to retrieve).
Why it matters — and what’s next
The model has knowledge and memory, but every tool/source is wired in ad hoc — is there a standard plug?
It can't — it's stateless. Memory is a system you design: what to save, the store it lives in, the retrieval that brings it back. Skip it, and every session starts blank. So a stateless model remembers only what you store for it: short-term in the window, long-term in an external store. Memory is engineered. You've given it knowledge and memory — but every tool is still wired in ad hoc. A standard plug? Next group: MCP.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
The model is stateless — it remembers nothing between calls. So when an agent seems to remember you across a session, or even next week, that memory isn't in the model. It's scaffolding you build around it, in tiers.
Tier one is short-term, or working memory. It's what's in the context window right now — the current conversation, recent steps, a scratchpad. Fast and immediate, but finite, and gone the moment the window clears. This is where context rot bites.
Tier two is long-term memory: an external store the agent writes important things to and retrieves from later — often a vector database, so it's semantic-searchable. This one persists across sessions. And retrieving from it is just RAG, run over the agent's own notes.
So next week it remembers you not because the model changed, but because it saved a note last time and retrieved it now. Write to the store on the way out; pull the relevant piece back on the way in. Memory is retrieval over a store the agent maintains.
People split it further — episodic memory of past interactions, semantic of facts it learned, procedural of how to do things. The names help, but the pattern underneath is always the same: decide what's worth saving, where to put it, and how to fetch it back.
The trap: expecting the model to just remember across sessions. It can't — it's stateless. Memory is a system you design: what to save, the store it lives in, the retrieval that brings it back. Skip it, and every session starts blank.
So a stateless model remembers only what you store for it: short-term in the window, long-term in an external store. Memory is engineered. You've given it knowledge and memory — but every tool is still wired in ad hoc. A standard plug? Next group: MCP.