toganTogan · for builders
Signals flow inand get distilled. An understanding of you sits in the middle as the source of truth. What reaches the model is never that store itself. It’s a compact projection of it, plus fetch-on-demand.
Capture: raw signals
Every turn appended to the episodic layer. Unconditional, cheap, no inference.
Materiality gate
Cheap heuristics: did this move, resolve, or contradict something already held? Most turns stop here.
Distill: server-side
Streaming and incremental. A cheap model tier climbs the new signal into typed nodes and links. Never a full re-climb.
The understanding
Typed entries connected by typed, directional, evidence-backed links. Every node carries confidence, weight, recency, and status. The source of truth. The model never reads it raw.
links: derived-from · serves · explains
Side by side
Docs you wrote: journal, identity, rules, projects. Ground truth. Irreplaceable.
→ build: highest-confidence signals
← evidence: nodes cite docs, not copy
Retrieve + rank
Select the relevant slice, rank by weight × confidence × relevance. Index-fetch discipline, over the graph.
The model-actionable payload
A compact projection of your understanding, assembled for this turn.
Values → the what (criteria) · Patterns + Preferences → the how (protocol)
The model, on any surface
Conditioned on the payload. Pulls specific facts, docs, and evidence on demand. Understanding primary, facts on demand.
Projections · views over the understanding
the graph is one view, not the only one
The one thing to hold
The model never touches the raw store.It’s involved in exactly two places, and neither reads the graph directly.
On the way in, a cheap model tier distills signals into the understanding. On the way out, the reasoning model is conditioned on a compact projection and fetches specifics only when a task needs them. The understanding is the asset. The projection is the interface to it.
Write path: building understanding
Every interaction on any surface, plus imports, connected sources, and your own docs, is appended as raw signal. Cheap and crash-proof.
Cheap heuristics decide whether the turn changed anything worth distilling. Most turns don’t, so most cost nothing beyond the append. That keeps streaming affordable.
When a turn is material, the engine climbs just the new signal into the understanding, with evidence and confidence. Server-side, so it works the same no matter which surface produced the turn.
Read path: how the model uses it
The relevant slice is ranked by weight × confidence × relevance.
Values give the model your criteria, patterns and preferences give it how to work with you. Scarce context spent on protocol, not a full profile.
The payload is the context. Beyond it, the model pulls specific facts and the evidence behind any claim only when a task needs them.