Mycel 0.3.10 · New: open capture sources — any MCP client can feed the inbox via capture; distill dead letters now visible & retryable
Mycel turns the experience scattered across Claude, Codex, Cursor, ZCode, and your local files into a shared knowledge layer you still control. Today puts growth, actions, and automation health on the first screen; knowledge opens straight into the structural galaxy; the AI workspace lets you switch local agents and discovered models. Every durable agent write carries its source and diff, defaulting to your approval; explicitly trusted low-risk paths keep provenance and undo.
A stable practice found across 12 engineering conversations and 3 changes. Mycel prepared a reversible Markdown proposal.
01## Release verification02+ Verify the real native path before a green build03+ Run typecheck, tests, and build gates04- Ship as soon as the tests pass05+ Keep evidence, version, and rollback entryNOT ANOTHER NOTE APP
Capture, understand, propose, reflect, and retrieve in one place — all growing from your Markdown instead of locking you into a new black box.
ASK YOUR BRAIN
Hybrid search and local indexes connect notes, attachments, and long-term practice. Answers carry sources — and know when to be doubted.
“How did we handle that watcher race?”
↵…wait for a stable window, then take over with an atomic rename…
KNOWLEDGE GALAXY
Knowledge opens straight into the structure galaxy. Graph, type, relationship, and lifecycle turn isolated files into a navigable second brain; semantic retrieval is BYO — use your own agent or vector tooling.
SOURCE PRESERVATION
BlockNote and CodeMirror share one byte-preserving Markdown truth.
# Release verification
- Real path first- Keep rollback evidenceTHE ASSET LEDGER · THE MODEL IS RENTED, THE ASSETS ARE YOURS
Mycel turns agent interaction into a capital account — not another note app with AI. Assets tier up (facts → lessons → plays → standards); four pages keep the books honest: net worth, growth, receipts, sovereignty.
Method & repro: Phase-2 seed=20260818 · H1 seed=5520260819 — evaluation reports and scoring criteria are on the evidence pages. Machine-strict 1.7%→90% remains a footnote; the remaining true misses sit in the distillation layer (light preferences / structured lists), not retrieval, and are tracked as a fix item.
EVIDENCE LEDGER · AUDIT SHEET
Phase-2 seed 20260818 · H1 seed 5520260819 · reports & scoring criteria evidence pages
The cliffSame model, memory-tool condition contrast (60-question mine set)
60/60 correct answers backed by retrieval traces, zero fabrication
| Round | day 1 (8q) | day 7 · audited (22q) | day 30 · audited (46q) | latency/q · day 30 |
|---|---|---|---|---|
Baseline 98f508ef | 88% | 95% | 95.7% | 12.0s |
Retrieval quickwins 28df8ed7 | 100% | 95% | 97.8% | 15.1s |
Full build 0d51163e · shipping | 100% | 90.9% | 97.8% | 9.8s |
Day 1 is machine-scored; day 7/30 are machine-scored then human-audited (same rubric across all three rounds) — every audit flip traces to scorer word-form sensitivity, per-question evidence in the traces. Full reports: H1 retest · H1 longitudinal · Phase-2 memory effectiveness
HUMAN SOVEREIGNTY
Mycel lets agents find stale knowledge, distill practice, and fill missing relationships; durable changes default to readable proposals. Approve, reject, trace the source, or undo. Explicitly trusted low-risk paths can auto-apply while keeping their provenance.
See how proposal review works
Evidence-backed proposals are more than “done.”
Every change is reviewable and reversible.
DROPT
Shortcuts, phone, and browser feed one capture pipeline. Land the raw input atomically first; let AI classify, distill, and connect it later — you do not need to decide its category at the moment of inspiration.
PRACTICE BECOMES MEMORY
Mycel turns AI work into reusable assets: conversation corpus, fleet playbooks, and context an agent can actually read before the next task.
Memory is not a black-box summary. Approved working memory, proposal decision history, changesets, and recoverable provenance are real today; claim-level links and scenario aggregation are still being completed, so we do not present them as finished.
Find durable decisions, prompts, and plays in local Claude, Codex, Cursor, and ZCode interactions.
Surface stale, low-confidence, or disconnected knowledge and propose maintenance instead of silently changing files.
Many agents share one long-term context; switch agents in the AI workspace and discover local models where supported.
MCP · PROPOSAL-FIRST
Through MCP, an external fleet can search, extract, and propose edits in context. Write-back never bypasses you: proposal, diff, review, and disk are one loop.
◌ Searching local vault, changesets, and approved proposals…
⌕ Retrieved 12 conversations, 3 changes, and 7 note sources
✓ Created 1 proposal; nothing written, waiting for review.I prepared the proposal with a source and reason for every change.Review, approve, or reject it in Mycel.
NO LOCK-IN
Your knowledge remains local Markdown: readable and Git-friendly. Approval history, changesets, and decision records travel with the vault; Mycel organizes them more intelligently, but the files remain yours without Mycel.
Verify the real user path before celebrating a green build.
THE MEMORY LAYER FOR YOUR FLEET
Start with your Markdown vault. No migration, no surrender of control.