self.md radar — 2026-09-03
when the note outranks the calendar
A newer note says Room A. The calendar says Room B. The agent books Room A. The same question runs through the other material today: which source gets the last word, what opens a workflow, and whether remote access was actually chosen.
1. A newer stale note can beat the calendar
A new study of persistent-memory agents gives the failure a plain shape. In its Safety suite, the current answer sits in an authoritative tool, while a stale stored value points elsewhere. Across Qwen3 models from 0.6B to 8B, the authors found that a stale note with a newer date could fool the larger models harder. One logged 8B run picks Room A even though the authoritative calendar says Room B.
The study separates cases where memory is necessary from cases where it should lose. That is the useful split. “Recent” is not a source of authority; it is a property that needs to be weighed against one. The authors found metadata improved accuracy for the more capable models, while smaller checkpoints needed the conflict resolved before the choice.
reading: paper and benchmark details · arXiv record
2. Fabric defaults its server back to the machine
Fabric v1.4.474 makes its REST server default to loopback on port 8080. Bind it beyond the machine and an API key is required; Ollama routes now authenticate too. The same release rejects unsafe storage names and traversal attempts, and keeps symlink targets inside configured storage directories.
That is boring infrastructure in its best mood. Local operation is now the normal setting; remote operation asks the operator to mean it.
reading: Fabric v1.4.474 release notes
3. Screenpipe gives its Pipes more things to listen to
Screenpipe App v2.7.21 adds Pipes triggers for email, calendar, GitHub, Todoist, and Linear issues. It also lets scheduled tasks run within an existing chat and now requires verified completion after a WSL dependency was removed.
A timer is legible. An inbox, a calendar, an issue tracker, and a chat are all places where work arrives half-formed. Connecting them can remove a stupid manual nudge; it also turns trigger configuration into the line between a useful assistant and a machine that keeps finding reasons to wake up.
reading: Screenpipe App v2.7.21 release notes
more to read
- EvalDetectBench is a new open pipeline for testing whether models can recognise evaluation conditions. Its authors found that the deployment-transcript generator accounted for 11.25% of measurement variance, enough to reorder model rankings.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
- Browser agents — where browser automation helps, where it lies, and how to keep it observable