Facts, decisions, patterns
Facts are stable truths ("the API uses cursor pagination"). Decisions record what you chose and why. Patterns capture how you like to work. Transcripts are never stored as memory.
Memory & knowledge
Memory that answers to you.
GPTree remembers like a careful colleague, not a surveillance camera: structured facts, decisions, and patterns instead of raw transcripts, every entry scoped, every entry inspectable. Here is the whole system, technically.
What gets saved
Three kinds. Three layers. Three scopes.
Every memory has a kind (what it is), a layer (how it behaves over time), and a scope (where it applies). That structure is why recall stays relevant instead of noisy.
Working, episodic, semantic
Working memory serves the current task. Episodic memory recalls what happened in past sessions. Semantic memory holds the distilled, durable knowledge that survives both.
Account, project, conversation
Account-scoped memory follows you everywhere. Project scope seals client work into its own room. Conversation scope stays put. Cross-project recall exists, and it is a setting, not a default surprise.
The Inspector
Every memory, on the record.
The Memory Inspector lists everything GPTree knows, by kind, layer, and scope. Edit an entry that drifted, pin one that matters, disable one temporarily, or delete it outright. You can also create memory deliberately: highlight any passage and pin it.
- See, edit, pin, disable, or delete any entry
- Pin any selection to memory, on purpose
- Extraction runs automatically; nothing hides from the list
What never gets saved
The boundaries are part of the design.
Temporary chats save no memories at all and delete themselves after 24 hours of inactivity. Project-scoped entries never leak across clients. And none of it trains anyone's models.
- Temporary chats: zero memory, gone in 24 hours
- Scope walls between projects, by architecture
- No training on your data, on any plan
Where this is going
From memory to the knowledge engine.
For enterprises, memory grows into governed knowledge: a system where the unit of truth is an immutable source span, and every answer carries its evidence, its access rights, and its historical validity. The founding architecture is frozen; the build is underway, with GPTree itself as customer number one. Presented, like everything here, through Branch, Compare, and Merge.
Source-anchored citations
Every claim traces to an immutable source span through citations durable enough to survive parser upgrades and re-rendering. A broken citation is a tracked reliability metric, never a silent failure.
Bitemporal truth and stale answers
When a fact was valid and when it was recorded are separate axes, so "what was true in March" and "what did we believe in March" both have answers. When a source changes, the answers that depended on it get flagged.
Knowledge spaces and promotion
Personal, team, and organization knowledge live in isolated spaces. Sharing is an explicit, logged promotion, so a private note never silently becomes company truth, and answers label the provenance class of everything they used.
A judgment pipeline
Draft, decompose into claims, verify each against its anchors, search for counterevidence, revise, calibrate. Contradictions are preserved and served, never resolved by omission, and abstaining is a first-class outcome.
Get started
Never re-explain yourself.
Start free in under a minute. No card required. Use built-in credits or bring your own keys.