Breadcrumb creator Justin (Inner Loop) claims the released local, encrypted Mac flight recorder — capturing screen, meetings, and AI transcripts into searchable memory with teachable rules surfaced through 30+ MCP tools — becomes a practical whole-workspace memory layer for coding agents across Claude Code, Codex, Cursor, and opencode; sustained external adoption resolves it.
state: seedheat: lowuncertainty: mediumconvergesscott: highagent-memory mcp local-firstJustinInner Loop
What is this?
Breadcrumb is a Mac app by Justin (Inner Loop), released via Show HN, presented as a local, encrypted flight recorder and context manager: it captures screen, meetings, and AI transcripts into searchable memory and exposes teachable rules to coding agents through 30+ MCP tools, targeting Claude Code, Codex, Cursor, and opencode. The supplied snippets do not cover Breadcrumb itself — they map the field it enters: at least seven local-first, MCP-exposed memory layers for exactly that agent list (agentmemory ~13k stars, pathmark, ai-memory, agentvault, memorix, nlm-memory, mnemo), all built by indexing agent session transcripts via hooks or on-disk session files, with the field visibly mature by mid-2026. None of the snippets describe screen or meeting capture; if Breadcrumb's whole-workspace claim holds as stated, that human-context axis — not transcript indexing — is its only differentiator, and nothing supplied confirms delivery or adoption.
Why it matters to Scott
A shipped product independently implements Scott's Attention Flight Recorder framework nearly to spec — passive capture of screen, meetings, and AI transcripts into searchable memory surfaced to coding agents — a dated third-party receipt for a position his canon already holds (the screen+meetings merge is his Two-Stream Timeline Merge pattern; the teachable-rules surface is second-order learning productized). It also stages a live field test of capture-vs-compilation — whole-workspace capture plus retrieval with no compilation layer, exactly the bet his 'capture was never the bottleneck' doctrine says fails — and its 30+ MCP tools put market adoption data directly against his code-first/attention-budget position, while productizing the transcript-memory substrate he hand-built in search and search-conversations. Caveat carried from grounding: nothing supplied confirms delivery or adoption, so the convergence holds only as claimed.
ip:framework.attention-flight-recorderip:concept.capture-vs-compilationip:concept.two-stream-timeline-mergeip:concept.second-order-learningip:framework.code-first-architecturedev:project.search-conversationsdev:project.searchradar:concept.agent-memoryradar:activity-frames-agent-memory-compilerradar:anarlog-private-meeting-memoryradar:openai-computer-history-memoryradar:engrim-local-cli-memoryradar:hindcast-claude-code-session-recoveryradar:concept.local-firstradar:concept.agent-observability
queries asked of Scott's wikis
- ambient raw capture vs curated writes — agent-maintained wiki and memory design tradeoff
- MCP server tool-count limits and agent tool-budget design
- shared cross-agent memory store for Claude Code, Codex, Cursor — local-first architecture
- always-on screen and meeting capture — privacy, encryption, consent tradeoffs
- teachable rules learned from usage surfaced into agent context
- flight recorder / session replay for debugging coding-agent runs
Measured heat
now 0 pts/hpeak 5 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 554h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p55 vs 1032 stories at the 336h mark (now 554h old) — ahead of aws-project-spend-limits (1.1x), behind all-your-agents-session-monitor (0.9x)
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-10-03T15:50:31Z
First crack in the zero-adoption wall — one anonymous 'I use this' — plus an unanswered third-party challenge (30+ MCP tools ≈ 85k chars of definitions stacked on Claude Code's built-ins) that operationalizes exactly the attention-budget risk at the case's core. Meaning shifts from 'pure creator claim' to 'shipped claim with a thin first adoption datum and an open practicality question'; with velocity flatlined (2.6th percentile, 0/h) the case stays cold until a real adoption line appears.
2026-10-01T19:04:36Z
origin walked (opencode/cheap-glm, conf 0.9): anchor hn.story.49924943 -> echo.blog.8a1ba17b97 by Justin Vincent (Innerloop)
2026-10-01T18:59:47Z
grounded: converges/high — A shipped product independently implements Scott's Attention Flight Recorder framework nearly to spec — passive capture of screen, meetings, and AI transcripts
2026-10-01T18:51:06Z
case created — First-party Show HN release with a distinct whole-workspace-capture claim (screen + meetings + transcripts, not just agent transcripts) uncovered by existing agent-memory cases.
Decision trace
- 10-05 21:47review_screenjev screen: no material development (noul=0.06)
- 10-04 01:50repriceFirst crack in the zero-adoption wall — one anonymous 'I use this' — plus an unanswered third-party challenge (30+ MCP tools ≈ 85k chars of definitions stacked on Claude Code's built-in
- 10-04 01:48review_screenFirst third-party report of actual usage ('jsh42': I use this) provides new external adoption evidence beyond the creator's own claims, and the tool-definition overhead question (~85,50
- 10-03 03:32sensor_dirtycomment_update
- 10-02 05:04promote_anchororigin walk conf 0.9
- 10-02 04:59groundA shipped product independently implements Scott's Attention Flight Recorder framework nearly to spec — passive capture of screen, meetings, and AI transcripts into searchable memory surfaced to
- 10-02 04:51createFirst-party Show HN release with a distinct whole-workspace-capture claim (screen + meetings + transcripts, not just agent transcripts) uncovered by existing agent-memory cases.