Engramlab: RL teaches models to use their notes, robust through millions of token edits
_AustinCalvert_ · x · 2026-11-06
Engramlab's new blog investigates how RL complements text-based memory by training models to actually use their notes. Key finding: training can be cheap compared to building the notes, and stays surprisingly robust even through millions of tokens of memory edits. Reposter lefttailguy frames it as "long ontology": local knowledge both makes models better at tasks and cheaper to teach new ones — the alpha is knowing what to write down, when to retrieve it, and crucially when and how to revise it.
Related event: Engram: RL Teaches Models to Use Their Notes, Cheaply and Robustly(11 posts)→
More from coding & agent
- Not Claude Code: the Claude model itself leaks your email from Cargo.toml, finds mitsuhiko — mitsuhiko · 2026-11-06
- Google's Antigravity coding app gets a sleek redesign, more updates teased within hours — clmt · 2026-11-06
- Perplexity cuts Agent API prices across effort levels, adds new ultra tier for long-running tasks — denisyarats · 2026-11-06
- Log failures to Datadog, let a bot fix them: an autoregressive loop trending toward 100% — msfeldstein · 2026-11-06
- Self-contained MCP server slips 59% of Opus outputs past Pangram AI detector at $0.30 per attack — amplifiedamp · 2026-11-06
- Google open-sources Agile Deliberation, an AI framework that turns fuzzy classification rules into explicit ones — aigclink · 2026-11-06