The model never remembers. Something around it does.

The chain of limitations — 6 sheets

1. 01Working Memory→The active context the model can see right now — the desk. Every token in the prompt is active state, and it runs on a budget.
   ↓

   Working memory breaks two ways: it ends when the session ends, and it fills up. Nothing survives the boundary on its own.

2. 02Episodic Memory→What the agent remembers happening — a specific investigation, a specific debugging session. The key property: it has a when.
   ↓

   The past session says what happened last time. It cannot say what is true right now — or whether last time still holds.

3. 03Semantic Memory→Standing facts that float free of time: repo layout, project rules, user preferences, name-to-ID mappings. The agent just knows.
   ↓

   Facts go stale and contradict. Without conflict handling, an old truth sits next to a new one as an equal — and memory becomes friction.

4. 04Procedural Memory→The how-to layer: runbooks, checklists, skills, tool schemas, orchestration code. Knowing how to ride the bike, not just that it has two wheels.
   ↓

   Durable stores are just a filing cabinet. None of it matters until the right slice reaches the context window at the right moment.

5. 05Context Assembly→The engineering problem at the heart of memory: retrieve, rank, and assemble slices of every store into the working context. The model only ever works at the desk.
   ↓

   A system that remembers everything eventually remembers too much. Perfect recall with no forgetting is the coworker who reads out their entire diary.

6. 06Forgetting & Conflicts→Memory hygiene: temporal decay, contradiction handling, compression, manual curation. Old memories can stay — they just shouldn't outrank what's true.
   ↓

   If you can answer the five questions — what's in working memory, which session matters, which facts are current, which workflow applies, what should be forgotten — you have an architecture. If not, you have scattered state.