LiMiT · Deterministic workspace

Conversation and shared learning

Learn supported facts from chat, correct memories, teach by demonstration and contribute reusable functions.

Personal conversation learning

Supported personal statements are learned privately in deterministic chat. This is literal pattern matching, not unrestricted semantic understanding. Use Framework to pause learning and review or undo changes. Explicit teaching remains available while automatic memory is paused.

My budget is 1000 EUR

Correct and follow up

A different value prompts confirmation unless you explicitly mark it as a correction. A correction changes only your own workspace. Ask what is my budget, then use it in a supported percentage calculation.

Actually my budget is 1500 EUR

Inspect the learning trail

The latest 50 events are available through the learning API; up to 500 are retained. Undo refuses to overwrite a memory edited later. Forget removes the fact, revision history and learning receipts; transcripts are separate.

learning history

Learn from demonstrations

Send four separate messages with distinct inputs. The first three select a candidate; the last validates it. Once accepted, write double 21 to invoke the learned function. New examples trigger bounded synthesis and regression checks.

When I say double 1, I mean 2
When I say double 2, I mean 4
When I say double 3, I mean 6
When I say double 4, I mean 8

Contribute to the shared library

Only explicitly supplied names and numeric examples enter the shared queue. Use non-personal sample numbers. Contributions from different users can combine into a candidate; contradictory examples block promotion. Private transcripts are not shared.

contribute: {"name":"double","consent":true,"examples":[{"input":1,"output":2},{"input":2,"output":4},{"input":3,"output":6},{"input":4,"output":8}]}

Shared activation and rollback

An operator inspects a candidate and approves its current digest with the local administration CLI. That version becomes callable by all workspaces. Supplied-example tests are not universal correctness proofs; operator review remains necessary. Existing versions survive failed candidates.

community status: double

Use a shared function

After approval, call the shared name explicitly. Community functions never silently replace a private function or core tool. Pinned versions are supported until the function is retired.

community run: {"name":"double","x":21}

What evolving means here

Learning is triggered by incoming statements, examples and explicit contributions. There is no idle background mind, unrestricted web learning, arbitrary code execution, automatic language acquisition, or claim of general intelligence.

Read the methods and limitations or try the examples in the workspace.