A 7B model with the right memory architecture can outperform a 70B baseline on multi-hop QA.
Mnemebrain is the infrastructure that makes that possible.
AI memory that turns experiences into beliefs — with full provenance.
Long-running agents accumulate thousands of episodes. Without consolidation, memory grows forever and retrieval becomes noisy. Mnemebrain compresses experience into beliefs — without losing evidence. Not a better vector store. A new memory primitive.
The right abstraction is neither. It is belief consolidation: compress repeated episodes into durable beliefs while keeping the evidence trail intact. Memory that compresses. Provenance preserved.
Three vegetarian food interactions. What other systems do with them — and what Mnemebrain does.
Source episodes are soft-archived, never hard-deleted. If a conflict arises later, they are rehydrated for a fresh consolidation cycle.
Most systems fake explainability by generating text. Mnemebrain explanations are structural — graph traversals, not prose.
Calling explain() on a semantic belief walks back through the belief graph: the claim, its extraction method, the confidence formula, and the exact source episodes that were consolidated to produce it.
When new evidence conflicts with the semantic belief, the system does not simply retract it. Source episodes are rehydrated, the new evidence is added, and the consolidation cycle runs again. Epistemic lineage is never lost.
Each claim is a node in a directed justification DAG with evidence, confidence, TruthState, and provenance. Not a plain fact — a belief with causal history and a revision path.
Idle-time background processing clusters episodic beliefs. A coherence gate validates before extraction. Structured predicate analysis promotes stable patterns to semantic beliefs. Source episodes stay linked via consolidated_from.
Local PageRank over a bounded subgraph — spreading activation through justification chains, not cosine ranking across the full embedding space. Auto-selects between vector and graph mode based on KG coverage.
Full architecture spec: belief revision (AGM), reconsolidation scheduler, GoalNodes, PolicyNodes, pattern separation, multi-agent CRDT replication. Read the full spec →
On knowledge tasks that require memory across sessions, a 7B model with Mnemebrain outperforms a 70B baseline. Architecture compounds what scale cannot.
| System | Knowledge tasks | Multi-hop QA |
|---|---|---|
| 7B model (baseline) | 41% | 33% |
| 7B + RAG | 49% | 38% |
| 70B model (baseline) | 55% | 47% |
| 7B + Mnemebrain | 63% | 58% |
Belief Memory Benchmark (BMB) v1 — multi-session knowledge retention tasks. Results reflect belief consolidation enabled with default Phase 5A settings.
brain.believe() to store episodic beliefs. The consolidation daemon runs in the background, clustering episodes and extracting semantic beliefs automatically.brain.ask() to query beliefs. Call brain.explain() to get the full provenance chain. Call brain.revise() when new evidence conflicts.brain.consolidate() to force a consolidation pass. Useful in batch pipelines or after ingesting a corpus. Or just let the daemon handle it.Open-source Python SDK. No sales call required. Star us on GitHub or pip install to get started.