Build Journal

The latest milestones from the lab. The gauge below computes itself from the status table — it cannot lie.

      ▾ Read the full build journal 
      ▴ Collapse the build journal
    

✦Retrieval learns to read in full: the search index now reads each memory's whole text, not just its summary — multi-memory recall climbs in both languages and recall reaches its ceiling. Only how memories are found moves; the model's answers are untouched.
✦The memory now sits an automated exam on every build — multi-memory reasoning, facts that changed over time, the honesty to say I don't know, scored in French and English against a fictional life. 848 tests.
✦The memory finds its own voice — talk to it hands-free: ask out loud, hear the answer, keep your risk profile current by voice. Spoken replies optional; everything stays on the machine.
✦The last written limit falls: a day forgotten before the tombstone registry existed can no longer be resurrected from an old archive. Limit #1 of the threat model, closed the way it was opened — in writing. 790 tests.
✦The Wardroom is feature-complete: divert a question from the exit door to your local model instead of the cloud — nothing redacted because nothing leaves. Search the whole chat vault too, decrypted in RAM, no index on disk.
✦The Wardroom opens — a fully confidential chat with your local model, for ideas that must never touch a cloud: each conversation sealed under its own key, forgetting it destroys the key. The badge stays lit: nothing leaves this computer.
✦The machine now tends itself: LAM raises its own Ollama — spawns it at boot, watches it, revives it. And Columbo extends his beat to voice: a cloud assistant on speakerphone is checked against your memory like anything typed.
✦The submersible IS the app now: the solar system retires and the cockpit ships — every subsystem a lit instrument, the vault an airlock. A session token now guards every sensitive route, and the recovery key rotates from the cockpit.
✦The product signs its work: Local AI Memory — one name across the app, the code and this site. What the lab built under a codename now carries the name that says exactly what it is.
✦The audit sweep closes: seven deep audits, every finding re-verified by an independent pass from Claude Fable 5, Anthropic's frontier model. Fifty-three of fifty-four fixed, zero critical left, one deferred and logged in the open. Test suite 571 → 711.
✦Forgetting keeps its word, all the way down: erase a day and its distilled facts, summaries and journal die with it. And the limits that can't be engineered away are written in an honest threat model — a privacy product that hides its limits protects no one.
✦Forgetting learns precision: erase one memory — not the whole day — and it is gone from search, from the feed, from the nightly replay. The eraser, not the shredder.
✦The memory faces its first exam: a fictional life of 242 memories through the real capture pipeline, sixty-nine gold questions, a deterministic score with a local judge reading alongside. Real numbers, never retouched, published as they fall.
✦Véga takes her post: the memory testament is real — bequeath the distilled essence of a life to someone you name, or have everything destroyed, decided while you are alive. Total opt-in: without a written testament, nothing ever happens.
✦The system turns its gaze on itself: six deep audits sweep every layer, findings fixed the same day — hardening before the audit, not after. Test suite 434 → 571 passing.
✦"Remind me next time Éloi comes up" — prospective memory is born: reminders attached to people and places instead of dates, fired the day the context returns. One spoken phrase is enough.
✦The memory learns lived time: every memory now carries the date of the event it tells, resolved at capture — while "last week" still means something — and anniversaries resurface by themselves, year for year.
✦Pénélope begins to weave: while the system sleeps it notices contradictions and open intentions, and dreams them back — always as a question, never as advice. Three cards a night at most; one tap each, and a "no" teaches it to stay quiet.
✦Columbo steps in: when a cloud AI gets your life wrong, your memory catches it in the act — the exact sentence that's off, right next to the memory that says otherwise. Validated end to end, judged 100% locally.
✦The Sentinel finds its voice: fifty rebuilt questions covering every corner of a life — and a golden rule, never write the secret itself.
✦The browser extension takes its first breath: a real claude.ai conversation lands in the local memory, captured quietly in the background.
✦The resonance field goes live: memories cool over time, resonate through the senses, and resurface at night while the system sleeps — proven end to end on a full test corpus.
✦The exit door opens: nothing leaves the memory without an exact preview and your approval — the fifth mechanism (interrogation log) comes alive. Five out of five.
✦The resonance layer lights up: memories are found by association (people, places, projects) and grow stronger each time you recall them.
✦Listening now captures BOTH sides of a voice conversation, and the cockpit takes on the Sanctum colors — a rain of golden glyphs that parts around the core.
✦Nothing personal is readable at rest anymore: journal, archives and the search index are all encrypted — the index lives only in RAM, gone when you lock.
✦The vault: one key per day, sealed under your master passphrase with a printable recovery key. Full API lockdown behind it.
✦First lab build: capture, search, grounded consultation and the initial vault.

Built in the Open

Where We Are Right Now

This page evolves with the build. Check back.

Component	Status	Notes
Backend API (FastAPI)	✅ Working	Local API — running; every memory subsystem runs on it
Frontend Interface (React / Vite)	◐ Prototype	Local interface — the submersible cockpit shipped July 10, every subsystem a lit instrument; public demo July 2026
AES-256 encryption — daily files	✅ Working	Each day sealed under its own key, unlocked by your master passphrase + printable recovery key — one compromised file exposes at most a single day
Ollama as memory gatekeeper	✅ Working	Small local models via Ollama (a free local AI runner) do the scoring and answering; the gatekeeper decides what — if anything — ever leaves your memory
The Guardian — emotional scoring	✅ Working	Assigns priority, emotion, valence, duration to every memory entry
The Big Sister — raw journal	✅ Working	Append-only timestamped log of everything that happens — encrypted with per-day keys
The Pitbull — sensitive data detector	✅ Working	Instant, fully offline flagging of sensitive data — runs locally, nothing leaves the machine
The Sentinel — personal risk evaluator	✅ Working	A 50-question personal questionnaire, fully rebuilt and editable — your own definition of "sensitive", with a built-in golden rule: never write the secret itself. The profile now lives encrypted at rest, like every other personal byte
The Consolidator — nightly memory	✅ Working	Like sleep: deduplicates, ages into encrypted archives, and distills the day's stable facts overnight — with provenance links to the source memories. It now runs by itself every night, catches up on skipped days, and a backup snapshot rides along
AutoResearch — automated research	◐ In Progress	Used internally during development to improve how memory is structured, scored and retrieved — feeds the build process, not the end user
Indexer + Scanner	✅ Working	Fast retrieval pipeline across all stored memories — now reading each memory's full text, not just its summary
The Five Mechanisms — research blueprint	✅ 5 of 5 built	All five are live in the lab build — surprise scoring, episodic→semantic distillation with provenance, canary memories, recall confirmation, and the interrogation log (every memory access and every outbound share is logged, encrypted) — full list on the Local AI Memory page
Voice assistant — hands-free	✅ Working	Talk to your memory without hands: ask a question out loud and hear the answer, or keep your personal-risk profile current by voice — yes, fix that, skip, and a silence ends your turn. Spoken replies optional (offline voice), text always works; closing the panel cuts mic, listening and loop at once
Voice capture — hybrid sessions	✅ Working	Records BOTH sides of a voice conversation with a cloud AI (your mic + the AI's voice) — transcribed locally, stored as memories + summary, audio deleted right after
Resonance layer — emotion-first recall	✅ Working	Finds memories by association (people, places, projects) and strengthens the ones you actually recall — emotion decides, the vectors fade as it learns
Resonance field — living memory dynamics	✅ Working	Memories are no longer static cards: activation cools over time, senses and associations carry recall from one memory to the next (a shared smell echoes strongest — the Proust effect), and a nightly replay resurfaces what matters — nothing is pushed, you consult the feed when you choose
Self-model — Pénélope, the narrative layer	✅ Working	Pénélope weaves while the system sleeps: at most three cards a night — two memories that contradict each other (both quotes shown), an intention left open too long, and a dream that is always a question, never advice. One tap answers a card, and a "no" teaches her to stay quiet on that pattern; your validated cards become the raw material of a future self-portrait — and a confirmed dream weaves its link for real: your yes strengthens the memory graph itself
Event time — the two clocks	✅ Working	Every memory carries two dates: when it was captured, and when the event actually happened — "last week" is resolved the moment it's written, while it still means something. Search can filter by lived time, and memories resurface on their own anniversary, year for year
Prospective memory — context reminders	✅ Working	"Remind me next time Éloi comes up" — one phrase at capture (works by voice) attaches a reminder to a person or place instead of the calendar; it fires as a quiet card the day the context returns. Time-based follow-ups are covered by open loops — this is the other half of human prospective memory
Memory testament — Véga	✅ Working	No longer a study — built. Your memory outlives you only if you choose: bequeath the distilled essence of a life to someone you name, or have everything destroyed, decided while alive. Total opt-in: without a written testament, nothing ever happens. The legacy is pre-woven and re-sealed at every nightly consolidation, and a quiet vigil can act even with the vault locked. No third party, ever
Speaker attribution — voice dialogues	○ To build	Your mic and the AI's voice on separate channels — transcripts become true dialogues: who said what
Cloud AI compatibility layer	◐ Prototype	The gatekeeper is live: nothing leaves your memory without an exact preview and your approval — sensitive parts redacted, every share logged. Browser extension for automatic capture comes next
Columbo — hallucination detector	✅ Working	Cloud AIs hallucinate — including about you. Columbo compares what they claim against your memory (the only ground truth about your life) and flags the exact sentence a memory contradicts — judged locally, in two passes, nothing leaves the machine. Paste any AI reply to check it; the browser badge arrives with the extension. Since July he walks the voice beat too — what an assistant says out loud on speakerphone is checked like anything typed
Le Carré — the confidential wardroom	✅ Working	A fully local chat with your model, outside the general memory: one key per conversation — forgetting it destroys the key. Divert a question from the exit door to the local model, anchor an exchange into memory by explicit gesture, search the whole chat vault decrypted in RAM only. The model in use is displayed at all times — nothing leaves the machine
LLM keeper — self-tending Ollama	✅ Working	LAM spawns, watches and revives its own local model server — a status light in the cockpit, one-click revive; no more silent scoring failures because the model runner wasn't up
Memory machine hardening	✅ Working	Memory lives on ONE machine — no synced copies, fewer attack surfaces. Surface shrunk: strict local-only access, data permissions locked down at every start, unlock and passphrase-change both protected against brute-force, every unlock logged. The vault now auto-locks after fifteen minutes of inactivity, refuses any outbound address that isn't your own machine, and its writes survive a crash mid-save
Mobile — the phone as its own memory machine	○ Under study — spec written	Decision: never a bridge between a phone and the memory machine. Instead, for those without a dedicated computer, the phone itself would BE the memory machine (Android) — same encryption, still one single device. Spec written; built when real demand shows
Master key management	✅ Working	Passphrase unlock, passphrase change + printable recovery key working — no cloud, nothing ever transmitted, on the single memory machine
Memory query interface	✅ Working	"What did I discuss about X?" — search with filters + answers grounded in your own memories, clickable citations; sensitive recall asks confirmation first
Import from cloud AI	○ To build	ChatGPT history export, Claude export, Gemini Takeout → converted into local memory
Browser extension	◐ Prototype	Passive capture of AI conversations directly from web interfaces — first real conversation captured and validated end to end; more sites and browsers being verified
One-command install script	◐ Prototype	“install.py” sets everything up (environment, Ollama, models) and “--check” diagnoses the install — validated on Linux; Windows/macOS and a non-technical version coming
Local encrypted backup	✅ Working	Verified snapshots (per-file fingerprint) rotating across multiple destinations — an unplugged external drive is simply skipped; restore tested end-to-end; also runs at nightly consolidation. Always encrypted, never decrypted to back up
Forgetting controls	✅ Working	Erase a chosen day — or a single memory: the eraser, not the shredder. Forgetting keeps its word all the way down: distilled facts, archived summaries and the day's journal die with it, and the app tells you how many backup snapshots still hold a copy until rotation clears them. Mark memories permanent so they never age out. You decide
Memory stats dashboard	◐ Prototype	A live orbital map of the whole system — agents, vault, archives, resonance — every subsystem visible at a glance, all local
Memory benchmark — sealed record	✅ Working	A five-phase exam: a fictional life ingested through the real capture pipeline, a simulated night, gold questions, deterministic scoring with a local judge reading alongside, and the record sealed. Real numbers, never retouched — and the benchmark lives in its own repository so anyone can re-run it
Threat model — honest limits	✅ Published	What the system protects against — and what it honestly can't: the exact scope of forgetting, what backups retain until rotation, what a trusted machine means. Written down in the open, ships with the code — a privacy product that hides its limits protects no one. Limit #1 fell on July 11: days forgotten before the tombstone registry existed are now seeded into it, once, from the audit journal
Miroir — the self-portrait	○ Under study — spec written	Your validated Pénélope cards become the raw material of a self-portrait — a mirror woven only from what you confirmed, never from what the system guessed. Spec written; built once enough verdicts exist to weave from
Testing phase	🔨 In progress	Each component tested independently — real conditions, real data, 848 automated tests passing
Analysis phase	○ Upcoming	Full system review — behaviors, edge cases, memory integrity
Final fixes	○ Upcoming	Corrections before security audit — nothing ships broken
Community security ideas	⟡ Open	Spotted a flaw, a gap, a better approach? Security ideas from the community are taken seriously — before the audit, not after
Community feedback — any concept	⟡ Open	Guardian scoring, Sentinel criteria, Consolidator logic — all open to improvement from the community
Community architecture proposals	⟡ Open	A better idea for any component? The design is not set in stone — community input shapes what gets built next
Community testing — open beta	○ Upcoming	Real users, real conditions — community testers invited before public release
Deep audit sweep — AI cross-checked	✅ Done	Seven layer-by-layer audits — from architecture to authentication — conducted and independently cross-checked with Claude Fable 5 (Anthropic). 54 findings: 53 fixed, 1 deferred & documented. The formal security audit below still comes at the end — this sweep is how we get there clean
Per-agent audit	○ Upcoming	Each agent audited individually — Guardian, Big Sister, Pitbull, Sentinel, Consolidator, Gatekeeper, Columbo, Pénélope
Security audit	○ Upcoming	Done at the very end — when all components are stable
Free code release — here & GitHub	○ Upcoming	Full source published here and on GitHub — read it, verify it, fork it

Evaluation Phase · Hardware Testing Planned

Hardware-Level Memory

Maximum privacy: air-gapped physical hardware. No internet. No exceptions.

Software encryption protects your data. Hardware isolation goes further — a dedicated device with no Wi-Fi, no Ethernet, no cloud — nothing leaves by wire or by air. The memory system and its agents run 24/7 on this machine. It listens in the background while you use Grok, ChatGPT, or any cloud AI: capturing context, storing it locally, flagging errors — all offline.

How it works alongside cloud AI

🧑

You

Your computer

→

☁️

Grok / GPT

Cloud AI

←

🎙

Mic & Camera

Local capture

→

🔒

Jetson Memory

Air-gapped · offline

Conversations captured locally via microphone — nothing touches the cloud. Agents store context, spot errors, flag sensitive data.

Tier 1 Nvidia Jetson Orin NX 16 GB ~$900 USD

Silent edge AI device the size of a credit card. Sits on your desk, runs memory agents 24/7, connected to a USB microphone, camera, and speaker. Fully offline.

CPU	8-core Arm Cortex-A78AE v8.2, 2 MB L2 + 4 MB L3
GPU	1024-core NVIDIA Ampere, 32 Tensor Cores
AI perf.	up to 157 TOPS (INT8)
RAM	16 GB LPDDR5, 102 GB/s
Storage	microSD + M.2 Key M NVMe SSD
Power	10–40 W configurable
I/O	USB-A 3.2 ×3, USB-C, HDMI 2.0, CSI-2, GbE, GPIO
OS	Ubuntu 22.04 (JetPack 6.x)
Size	100 × 79 × 21 mm
Internet	Disabled — air-gapped

Tier 2 Unitree Go2 Pro ~$2,700 USD

Quadruped robot running Linux. Your memory follows you around the room. Full AI processing onboard, zero cloud dependency. Wi-Fi disabled for air-gap operation.

Onboard AI	Jetson Orin NX 16 GB (built-in)
CPU	8-core Arm Cortex-A78AE + RISC-V MCU
Sensors	4D Lidar L1 · Intel RealSense D435i · 4× fisheye cam
DOF	12 (3 per leg)
Speed	up to 3.5 m/s
Payload	up to 3 kg (top rack)
Battery	~2.5–4 h typical use
Weight	15 kg
OS	Ubuntu 20.04 (ROS2)
Internet	Wi-Fi disabled — air-gapped

🔬

Evaluation phase — Hardware acquisition planned for November–December 2026. Goal: validate that a quantized local LLM (Ollama) runs reliably on Jetson, that memory agents process audio in real-time, and that the air-gap holds under real conditions. Test results will be published here. Community ideas on the hardware setup are welcome.

Local AI Memory — Build