Seedthink is a coordinated system of ten specialized components. Each one is responsible for a stage of the learning cycle — from routing the first question to distilling verified knowledge back into the model.
The stack
A router picks the cheapest path that can answer. The graph holds verified knowledge, memory holds context across four layers, verification guards what gets kept, and distillation folds it back into the model.
Seed Core
Router
Graph
Memory
Verify
Distill
Storage
The central intelligence coordinator. Seed Core understands user intent, plans reasoning paths, manages memory, triggers verification, and schedules learning. It does not need to be the largest model — its job is orchestration and growth.
Decides which model answers, which tools run, whether the web is searched, code executed, images generated, or knowledge stored. Dynamically chooses from GPT, Claude, Gemini, Grok, DeepSeek, future foundation models, and internal Seedthink models.
Seedthink connects to foundation models, live web and academic search, scientific repositories, simulators, calculators, coding engines, and image/video/audio generation. It does not try to store the internet — it learns how to access intelligence when needed.
Every answer is analyzed. The system extracts facts, concepts, relationships, procedures, and reasoning patterns — converting temporary answers into structured knowledge.
Nothing enters long-term memory without verification. Multi-model consensus, source validation, and confidence scoring ensure only verified knowledge is learned. This is how Seedthink prevents hallucinations from becoming permanent memory.
The true brain of Seedthink. A living graph storing facts, concepts, relationships, procedures, and domain knowledge across science, business, and the humanities — continuously evolving with every verified insight.
Working memory for active tasks. Episodic memory for important experiences. Semantic memory in the Knowledge Graph. Model memory compressed inside Seedthink's own growing models.
Humans forget. Seedthink should too. Every memory carries a confidence score, usage frequency, last verification, last access, and age. Useful memories strengthen; outdated ones weaken and are archived — keeping the system efficient and uncluttered.
Knowledge changes. Seedthink continuously re-checks stored information against new sources — scientific discoveries, software updates, business changes, world events — so the system stays alive rather than frozen.
Where Seedthink grows. Verified knowledge becomes training data. Training data updates Seedthink's own internal models. Over time, the system becomes increasingly capable using its own learned knowledge.
Seedthink ships with two parallel substrates: the platform's own global Seed (shared by every user) and your Plant-a-Seed (private, owned, $19.99 one-time). Both are instances of the same tiny-LLM pipeline — they only differ in scope, access control, and who pays for the compute.
Every chat turn and every Seed ingest URL is parsed by the Knowledge Extraction Engine. The assistant's reply (or the fetched document) is sent through a JSON-mode extractor prompt that returns atomic claims, each tagged with a topic and a 0–1 confidence score. Opinions, jokes, time-sensitive chatter, and personal data are dropped at this stage.
Each candidate is embedded with the platform's embedding model and re-checked against existing memory (cosine similarity over pgvector). Duplicates above 0.92 similarity merge and their confidence is bumped. New facts above the 0.82 verification threshold are marked verified; lower-confidence facts are stored but flagged for revalidation.
The trial Seed every account ships with is a 50-fact sandbox against the global memory. A purchased Plant-a-Seed creates a private namespace inside the same pgvector store, scoped byuser_idvia the match_facts_scopedRPC. Grounding queries only see facts you own — your Seed cannot leak into anyone else's, and the global Seed cannot read yours.
Verified facts are batched into LoRA-style fine-tuning shards. The shared Seed trains a tiny base model on the global verified corpus; each Plant-a-Seed trains a private adapter on top of that base using only its owner's verified facts. The result is a small, cheap model that already knows your domain before the router ever calls a frontier LLM — every prompt makes both the shared base and your private adapter measurably better.
Frontier labs scale parameters. Seedthink scales the verification loop. A frozen trillion-parameter model that re-runs the same weights against the same questions cannot become more reliable — only faster. A system that verifies, stores, revalidates, and distills every interaction into a sourced knowledge graph gets measurably more trustworthy on a fixed parameter budget.
VIMM, in this architecture, is not a moment when a giant model wakes up. It is the point at which the loop closes tightly enough that the tiny Seed model, grounded in its own verified graph, outperforms the frontier model it was bootstrapped from — measured by the Verified Intelligence Index (VII) and the Continuous Intelligence Benchmark (CIB).
This is the exact path a single user message takes through the stack — from the moment it leaves the browser to the moment a new verified fact is written into your Plant-a-Seed and queued for the next distillation run.
Browser POSTs the chat turn to /api/public/v1/chat with the user's session token.
api-gateway.server authenticates, applies rate limits, and logs the usage event.
embedText() vectorises the last user message; match_facts_scoped pulls related facts from your Seed scope in pgvector (threshold 0.7).
System prompt + grounded facts + chat history go to the Lovable AI gateway. Router picks the tiny model or the research model.
extractAndStore() runs a JSON-mode extractor over the assistant reply and produces atomic candidate facts.
Verifier keeps facts above 0.82 confidence; dedup merges anything above 0.92 cosine similarity with existing memory.
New rows are inserted into learning_events with your user_id, embedding, confidence, and verified flag.
The nightly job trains a fresh LoRA adapter from your shard; the router rebinds your Seed to the new checkpoint.
A Plant-a-Seed is not a new database, a new model, or a new endpoint. It is a scoped slice of the existing system, billed once and attached to your account forever.
Stripe webhook hits /api/public/webhooks/stripe, inserts a row into owned_seeds with youruser_id and bumps your monthly unit budget by +200. The row is the entitlement.
match_facts_scoped and the ingest writer both read owned_seeds before allowing inserts past the 50-fact trial cap. No owned Seed = 50-fact ceiling. N owned Seeds = unlimited fact storage on them.
The nightly distillation job groups verified facts byuser_id and trains a per-Seed LoRA adapter on top of the shared tiny base. Your adapter only ever sees your verified facts.
When the Intelligence Router picks the tiny model for a cheap turn, it loads the shared base plus your latest adapter checkpoint. Cold inference cost stays flat; quality grows with your verified corpus.
Watch one user turn move through the loop. The reasoner gets no smarter from this one prompt — but the verified knowledge graph does, and that difference compounds across millions of turns and shows up directly in the Verified Intelligence Index.
"What's the half-life of caesium-137 and why does it matter for soil remediation after Fukushima?"
Embed the prompt → match_facts_scopedreturns two prior facts from your Seed: "Fukushima Daiichi melted down March 2011" (0.94) and "caesium-134 has a 2-year half-life" (0.81). Both injected as grounding.
Router picks the research model (multi-domain, citation-heavy). Lovable AI gateway returns a 380-token answer covering the 30.17-year half-life, biological vs. physical clearance, and the potassium-displacement remediation strategy.
The JSON-mode extractor returns three atomic candidate facts:
All three clear the 0.82 verification threshold. None match an existing fact above 0.92 similarity, so three new rows land in learning_events withverified = true and youruser_id.
That night's job pulls the new rows, appends them to your per-Seed training shard, and trains an updated LoRA adapter. Adapter delta: ~0.4 MB. Your Seed now knows caesium chemistry without needing the research model to recall it next time.
Tomorrow you ask "is strontium-90 worse?". Router picks the tiny model. Tiny model + your updated adapter + grounded caesium facts answers correctly, in 90ms, at 1/40th the cost — without ever calling a frontier LLM. Multiply by every prompt, every user, every night. That is how verified intelligence compounds.
Every verified write from a planted Seed emits a sanitised delta — fact, topic, embedding, confidence, but never raw user content — onto an internal substrate stream. The shared Seed consumes that stream to widen its own coverage. The reverse channel pushes high-confidence shared facts into per-Seed candidate queues, where the user's own verification rules decide whether to admit them. Sovereignty is preserved at the boundary; verified intelligence compounds in both directions.
Each planted Seed also exposes its private knowledge through a per-Seed bearer-key endpoint — separate from the platform API — so users can call their personalised verified intelligence from any app they ship.
Papers, benchmarks, and open questions behind the Seedthink architecture — verified intelligence, compounding learning, and the road to AGI.
Explore Research →