2026-10-11 17:09 UTC

Independent use will determine whether Activity Frames can convert passively captured desktop activity into useful persistent agent memory without LLM-based compilation.

state: expiredheat: lowuncertainty: highconvergesscott: mediumagent-memory activity-capture memory-compilersnossa-y

What is this?

Activity Frames is a system described by Nossa Iyamu that deterministically compiles passively captured desktop activity into typed, bounded episodes for persistent computer-use-agent memory, without an LLM in the compilation loop. Its frames include application, site, timing, input-volume, and evidence pointers to raw capture data, making outputs byte-identical, cacheable, and mechanically auditable. The reported evaluation claims 98.4% accuracy over 128,756 frames spanning 51 active days, but the snippets indicate that all data came from one participant; they do not establish whether the approach remains useful or accurate under independent, multi-user use.

Why it matters to Scott

Activity Frames independently implements the core Attention Flight Recorder pattern—passive activity telemetry compiled into bounded, inspectable memory—and aligns with Cognitive Git’s separation of raw history, navigational representations, and evidence pointers. Its deterministic, no-LLM compiler is directly relevant to Scott’s dev-wiki and screen-extraction work and creates a dated-receipts/design-comparison opportunity, but the single-participant evaluation does not yet establish a result strong enough to change his position.
ip:framework.attention-flight-recorderip:framework.cognitive-gitip:concept.capture-vs-compilationdev:project.dev-wikidev:concept.progressive-screen-text-extractionradar:concept.agent-memoryradar:memory-bench-layer-baseline-validityradar:concept.computer-use
queries asked of Scott's wikis
  • deterministic memory compilers versus LLM summarization
  • passive activity capture as agent memory
  • agent memory auditability and provenance
  • episodic memory for computer-use agents
  • local-first desktop capture and privacy
  • evaluation of persistent agent memory across users

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟧 hnA compiler that turns a day at my computer into agent memory, no LLMnossa-y10
🟧 echo.github ⭐A compiler that turns a day of captured computer activity into agent memory without using an LLM.nossa-y——

Interpretation history

Decision trace