Repository navigation
Add agent memory benchmark - #362
Merged
Merged
Conversation
Keep Agent Memory Server history and files at the repository root. The public benchmark package lives in a subdirectory so the two trees can ship together. Co-authored-by: Cursor <cursoragent@cursor.com>
Landing copy now leads with the managed Iris product and points at the open-source foundation and benchmark without implying they replace it. Co-authored-by: Cursor <cursoragent@cursor.com>
The vendor classes lived in one providers.py file behind re-export stubs. Moving them into the named modules, with shared answering helpers in common.py, makes it easier to find and change a single provider. Co-authored-by: Cursor <cursoragent@cursor.com>
Neither exposes a Python API this harness can call, so the classes only ever raised. Carrying unrunnable providers, docs, and a test costs more than it explains. Co-authored-by: Cursor <cursoragent@cursor.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
tylerhutcherson
left a comment
tylerhutcherson
left a comment
Contributor
There was a problem hiding this comment.
Overall looks promising. A few of the cursor bug notes might be worth addressing. I also commented on some small readme improvements. Seems like we are still missing the redis implementation afaik
The public Iris REST API is now a first-class provider so runs can target Redis Agent Memory with an endpoint, store id, and API key. Co-authored-by: Cursor <cursoragent@cursor.com>
Vendor wrappers now match the locked Mem0, Vertex, Graphiti, Zep, and Oracle APIs. Ingest retries reset first, and per-example user ids cannot be overridden. Co-authored-by: Cursor <cursoragent@cursor.com>
Stop the judge from crashing on braces, avoid hanging LTM deletes, and clear AgentCore events between examples. Also put Redis first in the harness provider lists. Co-authored-by: Cursor <cursoragent@cursor.com>
Do not query until memory counts stop growing, write answers only after reset succeeds, re-download a truncated LongMemEval cache, and put Redis on the Apache appendix. Co-authored-by: Cursor <cursoragent@cursor.com>
Match the requested root preamble and make the Iris adapter the first registered provider. Co-authored-by: Cursor <cursoragent@cursor.com>
Stop treating a capped search page as extraction complete, walk Supermemory document pages, and stamp haystack dates on Zep, LangMem, and Vertex ingest. Co-authored-by: Cursor <cursoragent@cursor.com>
Wait on real record lists for Bedrock, Graphiti, and Oracle, and stop coercing store_id or project numbers from CLI params. Co-authored-by: Cursor <cursoragent@cursor.com>
Avoid holding the medium JSON in an HTTP buffer, delete Redis sessions by owner with pagination, and treat an unstable extraction wait as a timeout instead of a scored answer. Co-authored-by: Cursor <cursoragent@cursor.com>
…/agent-memory-server into add-agent-memory-benchmark
A failed bulk delete now raises so leftover owner memories cannot mix into the next ingest. Co-authored-by: Cursor <cursoragent@cursor.com>
Zep and Bedrock inherit a 30-minute wait instead of 120 seconds, and Mem0 ingest includes conversation dates so temporal questions can be answered. Co-authored-by: Cursor <cursoragent@cursor.com>
Bedrock, LangMem, and Supermemory were calling APIs with the wrong kwargs or a 100-item cap; resume also replaced wall time instead of adding it. Co-authored-by: Cursor <cursoragent@cursor.com>
Sync adapters were polling as if extraction were async, and Bedrock/Oracle dropped the conversation-date line other providers send. Co-authored-by: Cursor <cursoragent@cursor.com>
Those adapters extract asynchronously, so the runner must wait; other answering stores still return list_memories immediately. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Ruff wanted a shorter Redis delete path check and formatted three files; typos rejected the unparseable metrics key. Co-authored-by: Cursor <cursoragent@cursor.com>
Unscoped list plus dict equality missed Memory Bank rows, so later examples could see leftover memories. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.
Reviewed by Cursor Bugbot for commit af09623. Configure here.
The SDK cleanup path is delete_thread, so reset was a no-op when delete_user was missing and leftover memories leaked across examples. Co-authored-by: Cursor <cursoragent@cursor.com>
… file. Scoring had drifted to paraphrased prompts and a stricter yes/no parser; runs now also write the upstream question_id/hypothesis JSONL. Co-authored-by: Cursor <cursoragent@cursor.com>
State the benchmark as a best-effort sweep and keep the Remis research score separate from what Iris ships today. Co-authored-by: Cursor <cursoragent@cursor.com>
ichko
commented
Sep 15, 2026
tylerhutcherson
approved these changes
Sep 15, 2026
tylerhutcherson
left a comment
tylerhutcherson
left a comment
Contributor
There was a problem hiding this comment.
Looks good for this pass
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
Low Risk
Mostly additive docs and a standalone Python package with CI; production APIs are only touched when operators run benchmarks with credentials, and V0 server code is largely unchanged.
Overview
Introduces
agent-memory-benchmark/, an open LongMemEval v1 audit harness exposed as thememory-benchCLI (run,judge,providers). The shared loop ingests haystack sessions per example, waits for provider-specific extraction where needed, retrieves context, answers with a common prompt/model, and scores with official task-specific LLM judge prompts; runs are resumable and write auditable artifacts (answers.jsonl,hypotheses.jsonl,metadata.json, etc.) with sensitive params redacted.Nine provider adapters plug in behind a
MemoryStorecontract, including Redis Agent Memory (Iris REST API), Mem0, LangMem, Zep, Graphiti, Supermemory, Google Vertex Memory Bank, AWS Bedrock AgentCore, and Oracle Agent Memory, each with optionaluvextras and provider recipe docs.Repository landing docs are reframed for a two-project root (
V0/research server + benchmark harness), with README updates that center Redis Iris for production, explain the LongMemEval comparison and harness limits, and add citation links. GitHub Actions runs lint, format check, pytest, and package build for the new project underagent-memory-benchmark/.Reviewed by Cursor Bugbot for commit e25f7cc. Bugbot is set up for automated code reviews on this repo. Configure here.