The Dataset Viewer has been disabled on this dataset.

mcp-recall agent studies: evidence archive

The complete record behind six pre-registered studies of mcp-recall, a Claude Code plugin that replaces large tool outputs with summaries and stores the full text for retrieval. Each study asks whether that changes what a coding agent gets done, by running the same agent on the same task with and without the plugin.

This archive exists so the published numbers can be checked, not taken on trust. It holds every run the studies made, including smoke runs, superseded attempts and a void attempt. Those are labelled below and excluded from the analyses, as the reports say.

What the studies found

The findings, methods and limits are in three dated reports. Read those first; this card only says where their evidence is.

Report Studies Run folders
Does mcp-recall change whether an agent solves the task? 1: pilot on 45 hard SWE-bench Verified tasks, Haiku, Sonnet and Opus 5.5, plus a Haiku follow-up pilot-hard, pilot-heavy, followup-haiku
Does mcp-recall's compression change whether an agent solves the task? 2: the same tasks, three replicates, with the summarizing hook on and off comp-r1 … comp-r3
Does mcp-recall help in long sessions and with browser output? 3: long Django sessions. 4 and 5: browser tasks through Playwright MCP, before and after a bug fix. 6: the browser tasks on a self-hosted model long-s01 … long-s06, web-s01 … web-s03, web-fix-s01 … web-fix-s03, web-local-r01 … web-local-r03

In short, by each study's pre-registered reading rule: the pilot's harm rule fired for Haiku, but the report cannot tell that result from noise. Study 4 read as a cost, which turned out to measure a bug. Study 5, on the fixed build, read as a benefit at the edge of its rule. Studies 2, 3 and 6 showed no measurable difference. Agents almost never called the retrieval tools. The design and reading rules for every study were committed before its first run, in tasks/specs/414-task-success-claims.md; the commit hashes the reports cite are kept in the tag research/414-original-history.

Every folder in runs/

Folder What it is Counted in an analysis?
pilot-hard, pilot-heavy Study 1: hard set, and 10 of its tasks with a large-output prompt Yes
followup-haiku Study 1 follow-up: a second plain run and an install-without-hook arm Yes
comp-r1, comp-r2, comp-r3 Study 2's three replicates Yes
long-s01 … long-s06 Study 3: six chains of ten Django tasks Yes
web-s01 … web-s03 Study 4: twelve forum tasks, three sequences Yes (the report explains why its reading measured a bug)
web-fix-s01 … web-fix-s03 Study 5: Study 4 rerun on the fixed build Yes
web-local-r01 … web-local-r03 Study 6: the forum tasks as single sessions on a self-hosted model Yes
study6-infra-first-attempts Six Study 6 runs that failed for infrastructure and were rerun once, as the spec allows No, disclosed
void-study6-attempt1 Study 6's first attempt, voided before the restart No, disclosed
followup-check A one-task check that the follow-up's nohook arm was set up right No
smoke1, long-smoke-s01, web-smoke-s01, web-smoke2-s01, smoke6-r01, smoke6b-r01 … smoke6e-r01 Smoke rounds before each study No, disclosed in the spec
calib-full-output Study 2's prompt calibration rounds No
superseded/ The pilot's first launch, stopped because it kept no transcripts, and the first calibration round No
validate A harness check before the pilot No
pilot.log, followup.log, smoke1.log The pilot chain's console logs —

What one run contains

Each study folder holds a manifest.json (one entry per invocation: command line, prompt template, budgets, image digests, Claude Code and mcp-recall versions and SHA-256, harness commit, key spend), a run.log, and one record per (model, arm, task):

Path Contents
<model>/<arm>/<task>.json The run record: timing, exit code, the agent's final result event (cost, usage, turns) and what mcp-recall stored
<model>/<arm>/<task>.patch, .prompt.txt The graded patch and the exact prompt (SWE-bench studies)
<task>.evidence/stream.jsonl Every event of the session, tool results exactly as the agent received them
<task>.evidence/claude-home/, claude.json Claude Code's own transcript and config from inside the container
<task>.evidence/recall-store/ A snapshot of the mcp-recall SQLite store, full stored content included (recall and nohook arms)
logs/run_evaluation/ SWE-bench's per-instance grading logs
SHA256SUMS Checksums over every file in the study folder

The harness that wrote all of this is bench/agent/, and its README lists the harness's known limits.

Checking the archive

cd runs/pilot-hard && sha256sum -c --quiet SHA256SUMS

Every folder with a SHA256SUMS verifies this way. The archive has 26,678 files, about 3.2 GB.

What was redacted

This is a redacted copy. The original stays private as the audit record. Only infrastructure details were removed: a private gateway hostname, the self-hosted model's route name, the host machine's name and kernel string, and absolute paths. No credentials were found in the archive. Every change is listed in REDACTIONS.md. Because redacted files changed, the SHA256SUMS here were regenerated and will not match the original's.

Licenses and third-party content

See LICENSE.md. In brief: the material this project produced is released under CC BY 4.0. The archive also contains third-party content under its own terms: SWE-bench Verified task data and the source of the open-source repositories it draws on, MCPMark's browser tasks (Apache 2.0), and pages of the WebArena Reddit (Postmill) forum, whose posts derive from public Reddit content. The agents' outputs were generated with Anthropic's Claude models and, in Study 6, Qwen3-Coder-Next.

Citation

@misc{mcp_recall_studies_2026,
  title  = {mcp-recall agent studies: evidence archive},
  author = {sakebomb},
  year   = {2026},
  url    = {https://github.com/sakebomb/mcp-recall/tree/main/docs/research}
}
Downloads last month
48