memdebug 0.6.2

Added

• memdebug check and watch now say where an Open WebUI write came from. Until now that evidence (the app's own created_by label and how close a chat
  message was) was shown only in the viewer. For each added or edited memory, up to three are listed under the summary line ("where it came from: ... created_by=agent,
  type=user; ... a chat was active at the time"), with a line counting any more and a reminder that it is evidence, not proof. The text is cleaned for display like
  everything else memdebug shows. Other kinds of store describe no source and print nothing new.

Documented

• Open WebUI: a memory the model saves itself looks like one added by hand. Tried on a real installation: a memory saved by the model through Open WebUI's
  native add_memory tool carries the same created_by: manual label as one added in Settings, and in a Temporary Chat the timing line says "consistent with being
  added by hand". The label and the timing cannot separate the two (docs/agents.md, docs/threat-model.md). memdebug watch still records every add, edit and removal.
• Why an Open WebUI model may never save a memory, and an example router that fixes it (docs/open-webui-tools.md). A router (Pipe) function
  that rebuilds the request to Ollama drops the tools list, so the model can only pretend to save. The page describes what was seen and what was tried, and
  docs/examples/open-webui-smart-router.py is a router that forwards the tools. It is an example, not part of memdebug, and CI does not run it.

Fixed

• A git store with a very long history is no longer called quiet. memdebug reads at most 100,000 file changes of a git store's history, and at most
  100,000,000 characters of old and new text (every edit counts both versions, so a note of about 20 KB edited 2,500 times reaches the second limit first).
  Past either limit the reader stops, the newest history is the part that is missing, and the check for notes edited behind git's back is switched off: such an
  edit is neither flagged nor recorded. memdebug check still said "quiet, nothing new" with exit code 0, with only a warning after the summary. It now says
  "NOT READ IN FULL: ... the history is too long to read in full, so edits made outside git are NOT being checked" and exits 1. The gap itself remains: reading
  only the commits since the last look is not built yet. See docs/threat-model.md.

memdebug 0.6.1

Added

• Double-clicking the stand-alone Windows program now explains itself. Before, it printed its help and exited, so the console window opened and closed
  at once and nobody could read anything. Now, opened by double-click (a stand-alone build on Windows, no arguments, a console of its own, nothing redirected),
  it says that memdebug is a command-line program, shows three commands to type in PowerShell with the real file name, and waits for Enter. Started from a
  terminal, or with any argument, it behaves as before. The release smoke test starts the program like that on Windows and fails if it exits at once.

Documentation

• The Windows notes now describe the SmartScreen warning as it appears for a browser download of the stand-alone program ("Run anyway" / "Don't run" at
  once, publisher "Unknown publisher"), and say how to check the file first.

memdebug 0.6.0

Added

• Stand-alone programs. The release workflow now builds memdebug for Windows (x64), Linux (x64) and macOS (arm64) as a single file that needs no Python
  (PyInstaller), smoke-tests each on its own operating system, and attaches them to a GitHub release with a SHA256SUMS file and a build attestation. CI
  builds and tests them on every push. They are not code-signed (SmartScreen or Gatekeeper may warn), do not include Mem0, and still need git. See
  docs/releasing.md.

Fixed

• memdebug selftest in a build without its own Python: the tripwire checks could not start their tripwire and said SKIP, and "a hung process tree is killed"
  passed without testing anything (it started a program that exits at once). A stand-alone build now starts a hidden selftest-helper command with a fixed set
  of jobs instead. A source install is unchanged.