Skip to content
Memorie
Request access

The case for memory

Nine objections to agent memory, answered.

These are the arguments that come up in every evaluation call we take. None of them is silly. Most of them are right about something and wrong about the conclusion.

  1. 01

    Context windows keep getting bigger. Doesn't that solve it?

    It moves the failure, it does not remove it. Recall degrades well before the window fills, and the middle of a long transcript is the first thing a model stops reading carefully. A window is a buffer. Memory is a decision about what deserves to survive.

  2. 02

    Can't we just use a vector database?

    Similarity is not relevance. Embedding search happily returns the three turns that sound most like the question while missing the one line where the customer changed their mind. You need structure, and structure means entities and relations, not just chunks.

  3. 03

    What's wrong with writing transcripts to files?

    Files work until two agents write to the same one, or until you need to expire a fact, or correct it, or prove who read it. At that point you have written a database badly.

  4. 04

    Won't the model providers ship this?

    Some of them will, for their own surface. None of them will make it portable across the other assistants your team uses, and none of them will let you hold the keys.

  5. 05

    Isn't summarisation good enough?

    Summaries are lossy in exactly the wrong direction: they keep the narrative and drop the specifics, which is the opposite of what recall needs. Summarise for humans, structure for agents.

  6. 06

    How much does this actually save?

    The saving is in what you stop re-sending. Teams replaying full history on every call spend most of their token budget on text the model has already seen. Retrieval turns that into a bounded fetch.

  7. 07

    Doesn't a write step slow the agent down?

    Only if it sits in the request path. Ours does not. Writes go on a queue and get processed asynchronously, so the turn returns at the same speed it would have without memory.

  8. 08

    What stops the store filling with noise?

    Consolidation. Background cycles merge duplicates, promote what recurs across sessions and expire what has gone stale. A store with no forgetting policy degrades into a slower search index.

  9. 09

    Who can read our data?

    Contents are sealed with keys we do not hold. We can tell you how large your store is and how often it is queried. We cannot tell you what is in it.

Still unconvinced?

Send us a transcript that broke your agent and we will replay it against the pipeline.

Send a transcript