EunsuKim/swe-chat-memory-stores

SWE-chat Memory Stores

Every Claude Code memory note recoverable from SALT-NLP/SWE-chat, put back together from the tool traffic that carried it, and placed on the clock.

SWE-chat ships no memory files. What it ships is 5,851 session transcripts, and inside them the agent reads and writes files under ~/.claude/projects/<slug>/memory/. Those calls carry the file's bytes, so the store can be reconstructed — but only at the moments some session happened to touch it.

What is here

repos with a memory store	53 of 205
notes recovered in full	266 of 281
distinct observed states	452
observations	578
span	2026-02-10 .. 2026-04-19

How it was built

A note's text survives in exactly three forms, and only the first two are whole files:

evidence	gives
Read result	the entire file as of that moment, line-numbered
Write input	the entire file being written
Edit input	an old_string/new_string pair — a diff, never a snapshot

observations.jsonl is built from unbounded Reads and Writes only. Every row is a file that some transcript actually contained, at a timestamp that transcript actually carried. Nothing is interpolated between observations and nothing is carried across a gap, so no row is inferred.

Bounded reads (those passing offset or limit) are dropped, since a slice cannot stand in for a file. So are tool results that are notices rather than content, such as File unchanged since last read.

Fields

observations.jsonl, one row per observation:

field
repo	the SWE-chat repo the session belongs to
session	S01, S02, … ordered by the session log's own clock
ts	when the observation happened
note	the note's file name
path	its original absolute path, unmodified
via	read or write
sha	first 12 hex of the content's SHA-1
chars	content length
changed	whether the content differs from this note's previous observation
state_file	the matching file under stores/
text	the note, in full

Re-reads returning identical bytes share one state_file, so files count distinct states while rows count observations.

stores/<owner>__<repo>/ holds the same content as files: observations/ per timepoint, notes/ with each note's final text, and manifest.json with per-note recovery status.

Limits worth knowing

A note is only visible if a session touched it. Notes that existed but were never read or written leave no trace. Some are nonetheless named, because a captured MEMORY.md index links them — those appear in manifest.json as index_only, carrying the one-line index description and no body. There are 15 of them. Notes neither touched nor linked are invisible, and nothing here can bound how many there were.

Observations are sparse in time. A note observed on the 9th and again on the 15th was not necessarily unchanged in between; it was unobserved. Treating the series as a continuous state is an inference this dataset does not make. Measured on one store by predicting each read from the prior evidence, the prediction held 13% of the time wrong under a day of staleness and 100% wrong past a week — so the gaps are real, not incidental.

Content notice

The notes are reproduced verbatim from the upstream corpus, including operational detail their authors wrote for themselves: absolute home paths with usernames, host names and IP addresses, SSH invocations, and in at least one case a plaintext password. None of it was added here and none of it was removed; it is public in SWE-chat already. It is called out because extraction makes it searchable in a way that 5,851 transcript files do not.

Several stores also contain in-progress research — experiment status, reviewer critiques, acceptance estimates — for work that may be unpublished.

Attribution

Derived from SALT-NLP/SWE-chat, used under ODC-BY 1.0, and released under the same licence.

• SWE-chat: https://huggingface.co/datasets/SALT-NLP/SWE-chat
• Paper: https://arxiv.org/abs/2604.20779
• Snapshot: f66cca95b14caaa4177f7ed5eaa424608dadcffa

Rebuild it yourself with recover_memory_stores.py in this repo:

pip install huggingface_hub pyarrow
python recover_memory_stores.py --out rebuilt

Built 2026-10-13 with the extraction scripts in swe-multi (scripts/extract_swechat_memory_notes.py, scripts/dump_memory_observations.py).