A personal AI assistant that evolves with you. Memory, personality, proactive reach-outs — across macOS, Telegram, and Slack.
INTEGRATE — Pull this in now: it brings the most complete personal AI architecture in the vault, contributes harvestable patterns across memory, identity, and proactivity, and directly fills the declared gap for a richer Personal AI Memory comparison target.
19/24 — actively-maintained / well-documented / solid
Failed: H8: FAIL — README badges cover Docs, Discord, License, and "Built by Vellum" only; no CI badge (GitHub Actions, Travis, etc.) and no reference to .github/workflows/
Passed: H1: PASS — latest release v0.8.1 exists H2: PASS — v0.8.1 released 2026-05-13, same day as appraisal (well within 12 months) H3: PASS — last commit 2026-05-13, 0 days ago H4: PASS — last commit 2026-05-13, within 30 days H5: PASS — archived: false H6: PASS — 15 open issues, >0 and <100 indicating active triage H7: PASS — MIT license explicitly stated
Failed: D8: FAIL — no section headed "Limitations", "Caveats", "Known Issues", "Trade-offs", or "Not supported"; the note "CLI works but desktop app is our primary focus" is a caveat but not under a dedicated heading
Passed:
D1: PASS — README is present and extensive
D2: PASS — README well exceeds 1000 bytes; multi-section with tables, code blocks, and full feature descriptions
D3: PASS — explicit "Install" heading with bun install -g vellum and ./setup.sh from source
D4: PASS — code block under "Get started" and "CLI" headings with vellum hatch, vellum wake, etc.
D5: PASS — "Common commands" section documents CLI surface; documentation table lists "API & Communication" and "Commands"
D6: PASS — first lines read "A personal AI assistant that evolves with you. It learns how you work, remembers what matters, and acts before you ask."
D7: PASS — vellum.ai/docs linked multiple times in README with full documentation table
Failed: E2: FAIL — dependency manifest marked "Not available" in repo data E3: FAIL — cannot assess dependency count without manifest E4: FAIL — README does not mention test frameworks (jest, vitest, etc.) and no package.json available to confirm test script
Passed: E1: PASS — primary language is TypeScript E5: PASS — 389 stars, well above 50 E6: PASS — created 2026-02-07, ~3.2 months ago; ~121 stars/month far exceeds threshold of 2 E7: PASS — 52 forks, well above 5 E8: PASS — description is 113 characters, substantive and specific about memory/personality/channels
| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | Offers at least five harvestable architectural patterns directly relevant to PAI: SOUL.md + NOW.md for mutable personality files, fail-closed trust engine with single-resolution actor identity, hourly autonomous proactivity loop with interrupt suppression, manifest-driven sandboxed skills (SKILL.md + TOOLS.json), and hybrid dense+sparse retrieval with staleness windows per memory type. |
| Integration Readiness | 1 | This is a complete application (macOS app + CLI daemon), not an importable library; installing it as a companion process via bun install -g vellum is feasible, but extracting individual subsystems (e.g., the memory engine or trust layer) requires meaningful scoping work measured in hours rather than minutes. |
| Overlap Risk | 1 | Partial overlap with NorthwoodsSentinel--loam (Personal AI Memory substrate) and garrytan--gbrain (hybrid-RAG personal memory), but vellum-assistant goes substantially further with identity, proactivity, and multi-channel delivery, so it is not a near-duplicate of either. |
| Gap Fill | 2 | The landscape summary explicitly flags "Personal AI Memory: only NorthwoodsSentinel--loam — needs alternatives for comparison"; this repo is a feature-complete alternative and the highest-signal entrant seen in that category, directly addressing the declared comparison gap. |
Composite: 0.75
Category: Personal AI Assistant Crowding: 0 repos in vault (first-in-category) Alternatives: NorthwoodsSentinel--loam and garrytan--gbrain cover memory substrate overlap in Personal AI Memory; no prior Personal AI Assistant entrant exists vs. top alternative: garrytan--gbrain (standalone 4.3, PAI fit 5) provides hybrid-RAG memory and MCP tooling but has no identity layer, proactivity loop, multi-channel delivery, or sandboxed skill system Landscape impact: filling a gap — first complete personal assistant architecture in vault, elevates the Personal AI Memory category from infrastructure-only to application-layer
Density: 9/10 — Available: repo metadata (stars, forks, dates, license, topics), full README (8 KB), landscape context with prior appraisals, release data, crowding index. Missing: dependency manifest (package.json not provided), so E2/E3 could not be evaluated from direct evidence.
The commercial backing (Vellum AI, managed cloud offering) combined with MIT licensing creates an unusual sustainability posture — worth watching whether the open-source core stays feature-complete or gets hollowed out toward the managed tier. The SOUL.md / NOW.md / per-user journal pattern for assistant identity is the most novel primitive here and merits direct study independent of any integration decision. The lack of visible CI signals (H8) is the one unexplained gap for a repo at this maturity level.