Memory That Forgets

A persistent, self-revising hypothesis DAG for agentic R&D — exposed as an MCP server.

Current agent memory is passive: vector stores and scratchpads accumulate facts but never revise them. Hypotree structures the agent's working knowledge as a directed acyclic graph of hypotheses backed by SQLite-WAL. When an experiment fails, the engine walks the dependency edges and retracts what rested on it. When a premise collapses, every dependent subtree is pruned automatically.

---

What it does

• Write-back belief revision — an ATMS-style engine (de Kleer, 1986) that propagates evidence failures upstream through the dependency graph.
• Cascading prune — invalidating a parent hypothesis instantly transitions its entire subtree to PRUNED. No tokens spent on dead branches.
• Exclusion-group inference — confirming one member of a mutually exclusive group retires the rest as EXHAUSTED without probing them.
• Deduction by elimination — last-man-standing: when all but one alternative in an exclusion group are refuted, the survivor is VERIFIED without a probe.
• Backward pruning over a complete question — the dual of the above: when every candidate answer to a question is ruled out on its own evidence, nothing that assumes one of them can be satisfied, so those branches are PRUNED and the navigator names the question that ran out.
• The closed-world assumption is declared, not assumed. Both inferences above are sound only if the listed answers are all the answers. exclusion_closed=False says they are not — "which learning rate?" always admits another — and the engine then withholds both. And when a deduction it did draw turns out to rest on an incomplete list, it is withdrawn rather than defended: the node goes back on the frontier and one probe settles which premise was wrong.
• Thompson Sampling navigation — Beta-distribution sampling over the open frontier, giving bounded worst-case regret (no catastrophic lock-in).
• Conflict resolution via differential ablation — when an integration test fails but every component passes alone, the engine rebuilds the failing combination one swap at a time to pinpoint the culprit.
• A derivation trail, not just a state — generate_learning_path narrates what was settled, in order, separating what an experiment paid for from what the engine inferred for free, and calling out beliefs that were later withdrawn.
• Persistent across sessions, models, agents, users, and projects — the belief state is a SQLite database, not a context window.

---

Watch it think

A belief state that revises itself is hard to appreciate from a status column. The dashboard runs by default, beside the MCP server, so the graph is already there the first time you look for it: