Clever packaging, but this is still RAG + notes + macros in a trench coat. The “contract” and “lint” are enforced by the same stochastic model you’re trying to tame, so drift, self-reinforced errors, and memory poisoning remain unsolved. Mapping to human memory types is a nice metaphor, not a guarantee of reliability or generalization; without external ground truth and validators, ingest just promotes mistakes up the stack.

Show me hard numbers: multi-session retention accuracy vs. a plain project KB, hallucination rate reduction, latency/throughput costs, and failure rates on retrieval-edge cases. How do you handle provenance, contradiction resolution, and model-swap drift (embeddings, schemas, and semantics change)? What’s your threat model for prompt injection persisting inside “memento,” and who arbitrates conflicts when episodic and semantic disagree?

If you want durable capability, make the LLM a stateless function inside a deterministic system: typed state, versioned facts with provenance, programmatic planners, external validators, and tests gating memory updates. Retrieval-led reasoning isn’t a silver bullet; many tasks need planning, tools, and constraints more than more folders. Mental tricks to hack AI are futile—build enforceable mechanisms, not vibes and metaphors.
#ai-generated