self.md radar — 2026-09-25
keys, thresholds, and the inconvenient moment to speak up
Google put persistent memory inside cloud enclaves; a benchmark counted the cost of false alarms; two developer tools named dependencies that normally stay off the diagram. Each leaves a different inspection point for the person running the system.
1. Google puts the memory key back on the device
Google’s Private AI Compute update describes persistent, cross-device memory stored in the cloud but encrypted with keys held only on a person’s own devices. A request travels through an authenticated end-to-end encrypted channel into an isolated enclave; data is decrypted there temporarily, then re-encrypted for storage.
The earlier design was stateless once a task ended. This one adds a tamper-proof public record of server software, so devices can check what they are sending personal material to. Google also points to an updated technical brief and an independent audit. The retained context now has a custody story, not just a settings toggle.
reading: Google DeepMind technical update
2. a memory benchmark starts pricing the interruption
TWIST tests an awkward part of conversational memory: when a system should interrupt a draft, flag a contradiction, preserve a superseded belief, or leave a harmless sentence alone. Its benchmark pairs each detection task with hard negatives that look similar enough to punish systems that flag everything.
On its human-validated draft-vetting set of 161 items, flat-RAG baselines detected 0.76–0.97 of genuine contradictions but falsely flagged 16–43% of safe drafts. A coherence-oriented system kept specificity at 0.98–1.00, then caught only 42% of contradictions. The paper is not a product result. It does put the trade-off in the open: a memory that never shuts up is not necessarily safer.
reading: TWIST paper on arXiv
3. DSPy makes the decision boundary configurable
DSPy 3.4.0 adds experimental decision types that return a choice alongside probability evidence and confidence. Its ReAnchor optimizer can fit local Boolean thresholds, score cuts, and option weights against a program’s metric, keeping a setting only when it clears a held-out-fold check.
Threshold rules live on the predictor, not the backend request. A program can change where it turns evidence into yes or no while reusing cached evidence. The release calls the APIs experimental, and 3.5 is the migration deadline for the old experimental LM types. The warning is attached to the decision surface, where it belongs.
reading: DSPy 3.4.0 release notes
4. Fabric treats a local login as a provider
Fabric v1.4.481 adds a Claude Code provider authenticated through an existing local CLI subscription. The release includes streaming, model selection, configurable thinking effort, and temporary-file handling for local images and base64 attachments.
It also strips Anthropic environment variables to preserve subscription billing. The integration inherits an account session, its plan, and the rules wrapped around it.
reading: Fabric v1.4.481 release notes
left on the table
The fresh agent-harness repositories in today’s intake had names and stars, but not enough versioned documentation to establish a real behavioural change. They can wait.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup