A side-by-side teardown of the memory architectures most operators reach for, from a folder of markdown to a self-improving local-LLM stack. Measured on cost, velocity, time-to-insight, and what each one actually remembers when you stop typing.
| CAPABILITY | COL 1 · BASELINEClaude memory only~/.claude/projects/.../memory/ | COL 2 · NOTESObsidian alonevault + wiki-links + graph | COL 3 · HYBRIDObsidian + ClaudeMCP server + Smart Connections | COL 4 · OURSThe Volia Brainauto-grow · local LLM · RAG |
|---|---|---|---|---|
| Auto-grows from external sources | — | — | ||
| Local-LLM distillation (free) | — | — | — | |
| Vector semantic search | — | |||
| Decay / confidence tracking | — | — | — | |
| Contradiction detection | — | — | ||
| Cross-source consolidation | — | |||
| Voice briefings (MP3 / TTS) | — | — | — | |
| Auto-tweet drafts | — | — | — | |
| Newsletter / podcast drafts | — | — | — | |
| Spaced-repetition recall | — | — | — | |
| Self-quizzing / weak-spot detection | — | — | — | |
| Knowledge-graph visualization | — | |||
| Time-machine view of memory | — | |||
| Token cost ceiling | $0 | $0–10 | $40–150 | $5 flat |
| Setup time | 5 min | 1 hr | 1 day | ~2 weeks (one-time) |
| Maintenance burden | none | daily writing | tag + sync | reads itself |
A folder of markdown files Claude reads at the start of each session. Free, simple, and permanently stuck at "what you typed last week."
A beautiful linked-notes app with a graph view. Cheap, owned, durable. The catch: it is entirely passive. You are still the engine.
An Obsidian vault wired to Claude through MCP, with semantic search via the Smart Connections plugin. A real upgrade — and a meter that runs every time you ask.
An auto-growing memory system that ingests from nine sources, distills with local LLMs, tracks confidence, surfaces contradictions, and pushes back at you with briefs, drafts, and flashcards. The Anthropic bill is capped at roughly five dollars a month because everything heavy runs on your machine.
| Ollama | Local LLM runtime hosting qwen3-coder, deepseek-r1, qwen2.5-coder, nomic-embed. |
| launchd | macOS native scheduler for the five real-time file-watchers and morning-brief trigger. |
| cron | Twenty scheduled pipelines running on the host: ingest, distill, consolidate, publish. |
| Whisper.cpp | Offline speech-to-text for voice memos and podcasts. No upload, no API key. |
| yt-dlp | Pulls audio from YouTube and podcast feeds for downstream Whisper transcription. |
| ElevenLabs (Adam) | Premium voice render for the 7 AM morning brief MP3. Single paid touch-point. |
| Telegram Bot API | Push surface for briefs, alerts, flashcards. Inline-keyboard for recall scoring. |
| sqlite-vec | Embedded vector index. 768-dim, ~sub-100ms search across hundreds of thousands of chunks. |
| Vercel | Static hosting for voliaventures.com/brain dashboard and this report. Zero server cost. |
| Vault (markdown) | Obsidian-compatible plain text. Your notes are portable, greppable, and yours. |
--- name: 0dte_iron_fly_filter description: Filter to enter SPY 0DTE iron flies only when IV-rank is high and overnight gap is below 0.4 sigma. type: pattern confidence: 0.78 last_verified: 2026-04-22 source_url: - https://www.tastylive.com/shows/option-jive/episodes/0dte-mechanics - vault://memory/options_0dte_research.md decay_days: 45 contradicts: [] --- # 0DTE iron-fly entry filter Enter only when: - IV-rank > 35 on SPY - overnight ES gap < 0.4 sigma vs 20-day realised - VIX term-structure in contango (no kink at front month) Skip when: - FOMC, CPI, NFP, or large earnings within 6 hours - prior session closed at > 1.2 sigma move Backtest on file: 2018-01 to 2026-03, win-rate 71.4%, expectancy +$0.18 per dollar of max risk.
# Run with Ollama via subprocess; no API key, no network. PROMPT = """You are the Volia Brain distiller. Input is a long-form artifact (transcript, article, paper). Your job is to produce a memory note in this exact YAML+markdown shape: --- name: <snake_case slug, <= 6 words> description: <one-sentence claim, falsifiable> type: pattern | finding | hypothesis | decision confidence: <0.0 - 1.0, calibrated; default 0.6 unless explicit> last_verified: <today, ISO date> source_url: [<list of URLs and vault://paths>] decay_days: <estimate of how soon to re-verify> contradicts: [<slugs of any claims this challenges>] --- # <heading> 3-7 bullet points. Be concrete, name numbers, name dates. NEVER hedge with 'might' or 'could'. State the claim, then state the conditions under which it would be false. If the input contains a strategy: add a 'when to skip' list. If it cites data: include the dataset window and sample size. If it overlaps an existing memory, set 'contradicts' or merge. """ def distill(text): out = ollama.generate("qwen3-coder:30b", PROMPT + text) note = parse_yaml_md(out) embed = ollama.embed("nomic-embed-text", note.body) db.upsert(note, embed) classify_and_route(note) # deepseek-r1:8b decides library detect_contradictions(note) # claim-graph diff return note
A new authentication CVE drops at 03:14 UTC. The CVE feed pulls it within the hour, deepseek-r1 classifies it as high-severity for your stack, the contradiction detector flags it against an existing memory that says the affected library is safe. By 7 AM you get a Telegram message and a 30-second voice brief naming the file paths in your codebase that need to change.
A founder you follow drops a thread on a pricing experiment at 10:42. The Twitter watcher pulls it inside the hour, qwen3 distills the actionable claim, the brain notices it corroborates two earlier notes in your dashboards library and crosses the high-confidence threshold. A tweet draft lands in the queue with three sources cited. You publish it in two clicks before the take is cold.
Eight months ago you tried a momentum strategy and it lost money. Today an investor pitches a strategy that smells the same. You ask the brain. It returns the original kill-criterion note, the four trades that broke it, the post-mortem, and the date you decided to retire it. You decline the pitch in five minutes instead of re-running the experiment for two months.
A customer told you in February they hate weekend emails. You captured it in a voice memo on the drive home. Whisper transcribed it, the brain stored it with confidence 0.95 and a 365-day decay window. Six months later, on a Friday, you start to schedule a Saturday email blast. The brain pings: "this customer asked you not to." You change the segment and keep the relationship.
Most small businesses in 2026 are running on websites built in 2018, security gaps that would surprise their bank, and twenty hours a week of manual work an AI can do overnight. We do the whole job. The brain you just read about is one piece of a complete operational rebuild — and once we finish, your business runs on autopilot while you sleep.
A self-learning second brain for your business. Local-LLM driven. Costs a flat $5/mo to run after setup.
We replace your dated website with a fast, accessible, mobile-first site that converts. Same domain. No downtime.
OWASP-grade security review. We find what your competitors will be embarrassed by next year.
Show up in Google. Show up fast. Show up first when AI assistants are asked about your industry.
Replace twenty hours of weekly busywork with AI agents that run while you sleep.
Daily, weekly, monthly briefings tailored to your role. Voice-narrated. Sent to your phone.
Volia Ventures is one operator working out of Dallas, Texas. We build for ourselves first — every product on this list runs my own businesses today — and we sell what works. No retainer agencies. No sales calls. We do the work, you keep the keys.