Memory for AI agents
Are you interested in building agents that remember?
A memory system has to do three things right: tell your agent when it doesn't know, learn the moment you correct it, and keep recalling what it learned correctly as more memories pile up. No other memory system today can do any of these. Talos does all three.
We are opening early access to a small number of teams. If that is the kind of thing you want to build upon, tell us what you are building.
Why this is hard
Your agent keeps searching for what it already knows
Retrieval hands the agent the first few results and nothing else. When the answer sits below that cutoff, it looks exactly like an answer that was never stored, so the agent rephrases and tries again, carrying every returned document with it. Below is one run against a 511,962-document company corpus.
Asked: the quantization safety gate's default pass rate. One document in the corpus has it, and the agent has to guess the words that will return it.
The answer might be in there and need a better query, or it might never have been stored. The response is the same either way, and nothing tells the agent which. If rewording is guesswork, the agent is better off searching the corpus itself. On EnterpriseRAG-Bench, 90+% of all injected context was found to be irrelevant.
What memory saves
Teach it once. Then the search stops happening.
Same agent, same corpus, same questions every run. The only thing that changes is whether Talos recognizes the question.
What one correction does, across seven benchmarks
Each row is a public benchmark. Grey is what Talos recognized before; green is what one round of feedback added. Wherever it starts, it lands between 86.7% and 100%.
And it knows what it doesn't know
Each green dot is a benchmark, over 1,648 questions with no answer anywhere in the corpus. Search and the memory products built on it answer every one: a ranked retriever returns its top results whether or not any of them is the answer. The amber dot is a language model reading those candidates and throwing most of them out. That is the model’s judgment, not the memory knowing what it holds, and it costs a generation on every recall.
Turns per question, the agent’s default memory against talos
The same 20 BrowseComp-Plus questions
$1.98 vs $16.66
Talos against the agent’s default memory: 88% lower cost.
- −91% context re-reads
- 12.8M → 1.1M tokens.
Memory does not make search cheaper. It removes the search. Working the answer out took 15.4 turns and $16.66. Being told it took 2.5 turns and $1.98. The work of producing a memory and the work of using one are not the same work, and only the first one is expensive.
Tell us what you're building
Early access is rolling out in small batches. Everyone who writes in gets a reply from one of us.
Built by
Ram Sriharsha
PhD — ex-CTO Pinecone, Databricks, Yahoo Research
Ilshat Ahatov
ex-Pinecone, Meta, Microsoft