Grafy News
Try Grafy free

Jetriever Replaces Similarity Search with Utility Scoring for Agent Memory

A new retrieval method called Jetriever ranks memories by usefulness instead of similarity.

Illustration for the news story: Jetriever Replaces Similarity Search with Utility Scoring for Agent MemorySource: Grafy Research Lab

What is Jetriever and how does it differ from traditional memory retrieval?

Jetriever is a memory‑retrieval framework introduced by the Grafy Research Lab that replaces the classic similarity‑based lookup with a full‑pool utility scoring mechanism. Instead of asking which stored turn looks most like the current question, Jetriever asks which stored turn is most useful for answering the current question. The utility score, called Jev’s decision score, is computed for every entry in the memory pool, and the highest‑scoring entry is returned. This shift means that a useful memory does not need to share wording or surface features with the query – it only needs to provide the right evidence or context.

How was Jetriever evaluated and what were the results?

The authors tested Jetriever on the InMind benchmark, which contains 125 indirect‑query tasks where the answer must be drawn from a prior conversational turn rather than the query itself. According to the preprint, Jetriever placed the gold‑standard exchange first in 84.8 % of cases. By contrast, two strong baselines – BM25 and MiniLM – managed to put the gold exchange first in only 1.6 % and 0.0 %, respectively, when indexing the user turns.

A separate 20‑question comparison split the problem into two stages: finding evidence and ranking it. Both Jev (the utility scorer) and the large language model Qwen‑3.5 9B located 90 % of gold exchanges by rank 50, while the Laya model found only 30 %. When looking at the top‑ranked result, Jev succeeded in 85 % of the cases, whereas Qwen‑3.5 9B achieved 40 %. These figures highlight a dramatic advantage for utility‑driven retrieval over similarity‑driven baselines.

Why does utility‑based ranking improve indirect query performance?

Similarity search works well when the query and the relevant memory share lexical overlap. Indirect queries, however, often rephrase a problem, ask for an implication, or request a piece of evidence that was stated earlier in a different form. By scoring every memory for its decision relevance to the current question, Jetriever can surface a turn that directly resolves the query even if the wording diverges completely. The utility score incorporates the model’s own assessment of how the memory will affect the next reasoning step, effectively turning retrieval into a miniature decision‑making process.

What are the implications for AI agents and developers?

For developers building conversational assistants, retrieval‑augmented agents, or any system that must recall prior interactions, Jetriever suggests a new design pattern: treat the memory store as a decision space rather than a static index. Agents can allocate more compute to a single‑shot utility evaluation and still outperform traditional multi‑step similarity pipelines. This could reduce the engineering overhead of maintaining sophisticated embedding indexes while boosting performance on tasks that require nuanced context, such as troubleshooting, tutoring, or long‑form dialogue.

How to explore Jetriever‑style retrieval in Grafy

Grafy provides several tools that let creators experiment with utility‑based memory without writing a full research stack.

  • Grafy Chat lets you converse with multiple models side by side. By prompting a model to rank a list of previous turns for usefulness, you can manually emulate Jetriever’s decision score. Try the /chat interface and feed a short conversation history, then ask the model which turn would best answer a new question.

  • Grafy Graph visualises branching prompt‑output trees. You can store each turn as a node, attach a utility‑score attribute, and let the graph layout highlight the highest‑scoring path. This mirrors Jetriever’s full‑pool evaluation and makes it easy to compare utility scores against similarity scores generated by a separate embedding model. Open the /graph app to build such a view.

  • Grafy Reader is designed for research papers and PDFs. Load the Jetriever preprint, then ask the embedded model to retrieve the most useful citation for a specific claim. This hands‑on experiment showcases how utility scoring can improve literature search beyond keyword matching. Access it through /reader.

While Grafy does not yet ship a built‑in Jetriever engine, the combination of Chat, Graph, and Reader gives developers a low‑friction sandbox to prototype utility‑driven retrieval pipelines, test them on their own datasets, and iterate quickly.

Frequently asked questions

How does Jetriever compute the utility score for each memory entry?

Jetriever uses a decision‑oriented model (named Jev) that evaluates, for every stored turn, how likely that turn will help answer the current question. The model outputs a scalar decision score, and the turn with the highest score is returned as the most useful memory.

Does Jetriever completely discard similarity measures?

No. The framework still uses similarity as a baseline for comparison, but the primary retrieval step relies on utility scoring. In the reported experiments, similarity‑based baselines (BM25, MiniLM) performed far worse than the utility‑driven approach.

Can I use Jetriever’s technique with open‑source models in Grafy?

Yes. By prompting an open‑source model to rank a list of candidate memories for usefulness, you can approximate Jetriever’s utility scoring. Grafy Chat lets you experiment with different models side by side, and Grafy Graph can visualise the resulting rankings.


Source: Grafy Research Lab — Jetriever: Shifting Agent Memory Retrieval from Similarity to Utility