AgentMemory ShortTerm LongTerm Summarization
An LLM's context window is its only memory by default — anything not in it is gone. 'Memory' is the set of techniques for persisting/retrieving relevant information across calls despite that, and different memory types trade off differently.
Why we need this / what value this brings
Treating the context window as the only memory means every long-running or multi-step process is capped by how much fits in one call — real workflows (a 1000-device rollout) need information to persist and be retrievable beyond that.
When to use this
As soon as a workflow spans more than one LLM call and later steps need information from earlier ones that won't all fit in context together.
How to use or implement this
Pick memory type by what's actually needed: summarization memory for 'roughly what happened so far,' long-term/vector retrieval for 'look up this specific device's detail on demand,' and keep only the current step's essential context in short-term memory rather than accumulating everything indefinitely.
Research questions
- Short-term/working memory (what's in the current context window) vs long-term memory (stored externally — see AI/VectorStores — and retrieved when relevant) vs summarization memory (compressing older context into a running summary instead of keeping it verbatim): which does the chunk-stitching problem actually need?
- For the device-upgrade scenario specifically: does the final summary need memory of all 1000 devices' individual results (unrealistic to fit in one context window), or just an aggregate (counts, anomalies) plus the ability to look up any individual device's detail on demand — which is closer to long-term/retrieval memory than to keeping everything in short-term context?
- What's actually lost with summarization memory (compressing detail into a running summary) — the same information-loss question as chunk-stitching, but across turns/time instead of across chunks of one input.
Empty folder — drop notes, links, and findings here as you research.