Persistent per-user memory LIVE
Every conversation is analysed for facts worth keeping, and the right ones are quietly injected next time. Your AI stops re-introducing itself.
Free Developer and upEverything in the box.
All 54 of them, grouped by the problem they solve. Nothing on this page is a roadmap item — but plenty of it ships switched off, and every entry says which.
LIVE on by default
ONE SWITCH AWAY ships off, you turn it on
SHADOW measuring itself before it takes over
EXPERIMENTAL real, with rough edges
Problem — your AI forgets you every time
8 features
The memory engine
Three-lane fact extraction LIVE
320 hand-tuned regex extractors plus spaCy catch facts in microseconds, and a rate-capped local-LLM pass picks up what they miss. All three lanes run on your hardware.
Free Developer and upHybrid retrieval, ten categories LIVE
Keyword search and semantic embeddings run in parallel and merge through Reciprocal Rank Fusion, with weights that shift by intent. Identity, health, work, family, technical, preferences and four more categories each carry their own priority and expiry.
Free Developer and upAtlas — the belief engine SHADOW
Where the fast lane stores what you said, Atlas tracks what it believes about you: confidence scores, true-then versus true-now eras, mood-aware hedging, clarifying questions when you contradict yourself, and overnight adjudication by your own local model. In its evaluation new-vocabulary recall went 0.122 → 0.551 at 0.864 precision across 658 cases — and it still runs in measure-only mode until it earns promotion on your data. Ask it why with GET /api/atlas/belief/{id}.
Muse — the injection brain SHADOW
Injection grows a brain: expired and contradicted memories are guarded out, candidates are rescored by belief confidence, and a working set stops facts flickering in and out between turns. At most one gentle nudge per conversation. Measuring itself against the current behaviour first.
All tiersThe Gardener and the trust loop LIVE
An idle-time loop dedupes, resolves contradictions and keeps the corpus lean — never touching anything you pinned. Rate any injected memory up or down from the chip, and the ones that keep earning a thumbs-down fade out of rotation.
All tiersYou can see exactly what it knows LIVE
Every streamed reply carries a “N memories used” chip — click it to see what the model was told, rate it, or forget it. /my-memory is a phone-friendly page where each person edits, pins and deletes their own memories without admin access.
Knowledge graph and REM consolidation LIVE
Entities and relations land in a temporal graph where a two-hop walk surfaces memories neither search would find. While you are away the proxy consolidates: merges duplicates, links facts far apart in time, forms hypotheses, and greets you with an overnight digest.
Premium
Problem — your AI can’t read your documents
3 features
Files it actually understands
Upload and just ask LIVE
33+ file types, and UPtrim works out which file you meant. Ask about “employee compensation” and it finds the salary breakdown section without a single shared keyword.
All tiersThe Inspector LIVE
Thirteen format-specific analyzers build a structured dossier for every upload — and only run while the GPU is idle, so uploading never slows down the chat you are in.
All tiersBig files, private folders LIVE
Chunked, resumable uploads with per-user folders and opt-in folder sharing. Per-file ceilings run from 15 MB on the free tier up to 100 GB on Business.
All tiers
Problem — one model can’t do everything
7 features
Routing, scheduling and scale
Every backend in one dropdown LIVE
Models from all your LLM servers appear in one list, each with its own context window — Ollama, llama.cpp, LM Studio, vLLM, SGLang or anything OpenAI-compatible. vLLM and SGLang are detected and promoted with zero configuration.
1 backend free · up to unlimitedAliases and the Auto model LIVE
Users ask for coder and UPtrim translates it to whatever is actually serving that job today. The virtual Auto model routes each turn by task complexity.
A real generation scheduler ONE SWITCH AWAY
A proper queue per backend with priority classes and per-user fairness, so people see “queued #2 · ~40 s · preparing context…” instead of a dead spinner — and research keeps running while they wait. An overflow burst pool ships empty for you to fill.
All tiersScale profiles LIVE
home, team and business size ten interacting caps in a single move — 40, 150 or 400 concurrent requests — and are chosen from your license automatically.
Adaptive demand policy SHADOW
Under load everyone’s budget shrinks slightly so that everyone stays served, with floors that never drop. Currently measuring what it would have done before it does it.
All tiersFour cloud providers, opt-in ONE SWITCH AWAY
Anthropic (OAuth, no API key to paste), OpenAI, Google Gemini and OpenRouter’s 100+ models, with every call written to a cost ledger. Local stays the default, and local_only mode never leaves your machine.
The Conductor EXPERIMENTAL
Breaks one turn into typed subtasks, routes each to the best backend under cost, latency and quality constraints, runs them in parallel and assembles the answer. Ships off; treat it as experimental.
Pro and up
Problem — shared AI means shared memories
5 features
People, households and teams
Complete per-user isolation LIVE
Five-tier identity resolution decides who is talking before anything is stored or injected. User A never sees user B’s data, on the same box, in the same chat app.
All tiers · unlimited usersAccounts, invites and team pools LIVE
Self-registration with invite codes, account linking, and shared team memory pools with roles for the facts a whole team should have.
All tiersBudgets and curfews LIVE
Per-user daily, weekly and monthly token budgets, plus allowed hours per weekday in the right timezone. When someone hits a limit, the message says exactly that instead of failing weirdly.
All tiersParental controls LIVE
Child accounts with their own rules. Kids ask for more time from inside the chat; parents approve with a PIN that is never displayed, compared in constant time and checked on the server.
All tiersMulti-node fleet ONE SWITCH AWAY
Several proxies, one memory: the master owns the shared database, workers funnel every read and write to it, and users and settings replicate down. A master outage degrades to stale-serve and then to an honest failure — never a silent fork. Two honest notes: every node must run the same build, and the link between them is plain HTTP unless you terminate TLS yourself.
Business / Enterprise · talk to us
Problem — your AI doesn’t know anything recent
5 features
Ghost — live web research
The research engine LIVE
While your request is being prepared, specialist workers race a 5,000 ms deadline — fetching pasted URLs, searching the live web, deep-scanning your memories. Whatever finishes inside the window gets injected, so there is no added latency. Anything good that missed the budget is carried into the next relevant turn.
Standard and upSix search backend families LIVE
Keyless out of the box and pluggable when one of them stops working from your network, with results fused by a mini-RRF and each engine behind its own circuit breaker.
Standard and upDistilled answers, not link dumps LIVE
Fetched pages are condensed by a local model into four to six attributed claim lines above the raw sources. “I need more detail” re-searches the topic you are already on — and a genuinely new question never inherits the old one.
Standard and upSignal-scoring specialist router ONE SWITCH AWAY
A learned router scores all thirteen specialists from fifteen signals per turn, with per-route weights that improve from real outcomes. It ships in legacy mode — run it in shadow, then promote it.
Standard and upReflex search — mid-stream rescue ONE SWITCH AWAY
If the model starts to say “I don’t have access to current information”, Ghost catches it before a single byte reaches you, fires a search and regenerates the answer. Ships off; roll it out when you trust it.
Standard and up
Problem — long chats go off the rails
4 features
Context that holds
Running summaries and compression LIVE
The thread is summarised as it grows and compressed proactively at 60% of the limit, instead of falling over at the edge of the window.
All tiersThinking, rendered properly LIVE
Reasoning blocks are redirected into reasoning_content so your UI can render collapsible thinking — Qwen3, QwQ, DeepSeek-R and GLM-4 included — and untagged chains of thought get caught too.
Smart thinking budgets LIVE
Reasoning gets a budget per tier, and is switched off entirely for the kinds of turns where it only burns tokens.
All tiersVision and audio passthrough LIVE
Image and audio turns reroute to a capable backend automatically, older images collapse to placeholders so a long thread survives, and audio is transcribed when nothing can hear it.
All tiers
Opt-in territory — nothing here runs unless you say so
8 features
Agents and deliberation
Agent mode ONE SWITCH AWAY
Plans, acts and iterates — web search, file editing, code execution — across seventeen tools with hard caps and a per-tool on/off switch. Ships off.
Standard and upA sandbox that fails closed LIVE
File and shell tools are confined to a workspace unless you deliberately configure a project root. Forgetting to configure something makes the agent more restricted, never less.
All tiers, whenever the agent runsSub-agents and verified swarm ONE SWITCH AWAY
Sub-agents run subtasks in parallel, and Hyper Mode can dispatch them to idle boxes. The verified swarm pulls from a shared frontier and puts every claim in front of skeptic judges that refute by default, so only majority-confirmed findings come back.
Pro and upThe learning layer SHADOW
An eleven-signal scorer decides when a request is really agent work, and successful runs are distilled into reusable playbooks with outcome-weighted trust. Both are measuring themselves before they take over from the simpler behaviour.
All tiersThe experimental shelf EXPERIMENTAL
Speculative branching that snapshots files, tries a hypothesis and rolls back cleanly; read-ahead prefetch; headless browser embodiment that works with any model because it reads text snapshots; standing “watch X, tell me when Y” directives; and per-run token and wall-clock budgets. All real, all off or badged experimental.
Varies · see the config referenceEvolve — nightly self-review LIVE
Once a night the agent grades its own week and recommends config changes. It never applies them itself, and it says nothing when there is nothing worth saying.
All tiersMind Mode and Parliament EXPERIMENTAL
Ask for the virtual mind model and two or three of your models reason together — critique, verify, diverge. Parliament clusters your memories into topical personas and gives decision-shaped questions a short deliberation transcript. Neither fires unless you ask for it.
Predictive context ONE SWITCH AWAY
Pre-loads the context you are most likely to need before you ask for it.
Premium
For builders and automators
4 features
Surfaces you can build on
MCP server LIVE
A full MCP server at /mcp/ with 40 tools — memory, files and shell, LLM helpers, ten LSP code-intelligence tools, scratchpad, cloud delegation, blueprint tools — for n8n, Claude Desktop or any MCP client. Reads are auto-allowed; sensitive writes wait for approval in the dashboard.
Visualcoding — the blueprint IDE LIVE
Wire blocks pin-to-pin on a node board at /visualcoding and get a real, readable Python file back. The board and the code are the same artifact, round-tripped through the AST without losing anything. Nothing executes inside — it exports.
TrimScript plugin engine ONE SWITCH AWAY
A plugin language plus a native Python loader for extraction rules, injections, filters, middleware and agents — with fail-closed permissions, per-plugin state, versioned migrations and SSRF-guarded HTTP. Off by default, and still badged experimental.
All tiersuptrim-cli — the terminal app LIVE
npm i -g uptrim-cli puts a real coding agent in your terminal: inline diff approval, /agentic loops, /fix-tests until the suite is green, ripgrep search, a /dual split-screen review, and memory that follows you between sessions. Honest caveat: the step cap is the reliable brake — the dollar cap is best-effort.
How you reach it
5 features
Surfaces and access
The dashboard LIVE
Easy mode for the handful of settings most people touch, Expert mode for all 500+ — grouped, searchable and applied live — plus a guided tour and a Brain panel that maps what is actually switched on right now.
All tiersSix ways to chat LIVE
A browser chat SPA, a Python terminal client with live tokens per second and six themes, a native desktop window, and the chat apps you already use. The boot-time TUI monitor adds live status, backends, users and logs on Premium.
All tiers · TUI monitor PremiumOne-click tunnels LIVE
A Cloudflare quick tunnel gives you an instant public URL, or sign in and use your own domain with the DNS route created for you. Tailscale publishes at your tailnet name, and Funnel makes it public.
Standard and upVoice and image addons LIVE
Voice runs local Whisper speech-to-text with Kokoro text-to-speech, and Photo Gen runs Stable Diffusion. Both register as capabilities the moment you start them.
All tiersThe virtual pet ONE SWITCH AWAY
A pet hatches per user, grows with usage, and reacts to what the proxy’s brain is doing. Off by default. Yes, really.
All tiers
The boring parts that decide whether you keep it
5 features
Platform quality
Typed configuration, generated docs LIVE
Every one of the 500+ settings is typed, validated and clamped when you save it, with stability badges that mark what is stable, experimental or inert. The reference document is generated from the shipped defaults, so it cannot drift from reality. Nearly every setting applies live; the rare exception wears a RESTART badge.
All tiersSecurity hardening LIVE
A mandatory salted-PBKDF2 dashboard password even from localhost, a fail-closed sandbox, secrets masked in every API response, forwarded-header spoofing checks, DNS-pinned and SSRF-filtered outbound fetching, verified TLS and constant-time PIN comparison. A full adversarial review was run against it and every confirmed finding was fixed.
All tiersCredential scrubbing LIVE
Paste a key by accident and it never reaches the database: 36 built-in patterns catch credentials before storage — always on, no setting to forget. AWS, Anthropic, OpenAI, GCP, Azure, GitHub, GitLab, Slack, Discord, Stripe, npm, PyPI, Docker, Sentry, Datadog, New Relic, PEM private keys and JWTs are all covered. In shared or team memory the fact is rejected outright; an auto-extracted fact becomes a redacted placeholder that is never injected into a prompt; a manual save is redacted in place. Every catch surfaces as a brain event in the dashboard.
All tiersSkribe write engine SHADOW
One persistent writer, group commits, write coalescing and a separate audit database, crash-tested for turn atomicity. A two-second boot probe measures real fsync latency, spots hostile locations like cloud-synced folders and network shares, and warns loudly instead of silently corrupting.
All tiersIt looks after itself LIVE
A per-subsystem watchdog restarts what breaks with exponential backoff and quarantines what keeps breaking, backup and restore are one click, and the whole thing runs from one relocatable data folder on Windows, macOS and Linux.
All tiers
Coming soon
Two things are built but deliberately not switched on yet: the retrieval front door, which adds a chunk-level index so answers can cite “filename, characters X–Y”, and the unified maintenance engine, which folds every background cleanup job behind a single mode switch. Both ship dormant today. When they are on by default they will appear above with a LIVE chip, and not one minute earlier.
Wondering how young all this is? Read the early-access status — no testimonials, just dates.