- Current research, not archived coursework
This page traces David's current Agent_Memory research rather than a past class project. Every temporal and retrieval state comes from one deterministic run of the Python prototype; the browser reads that committed trace instead of reimplementing the memory system.
- Why flat RAG is only a baseline
A similarity search can find top-k snippets, but it does not by itself say whether a fact is current, trusted, permitted, or connected to the evidence that produced it. The first chapter is a comparison baseline, not functionality claimed for Memory OS v0.
- A write gate makes evidence governable
The gate explains each admission decision. It keeps a poisoned web observation as quarantined evidence while accepting instruction-shaped text only when it is verified/private policy—showing a trust boundary rather than a blanket command filter.
- A correction changes what is live
The detailed-summary correction closes the concise preference's validity window. Both records remain auditable, but only the later one can answer a later question; a deletion cascade prevents the original from silently coming back.
- Context follows the temporal entity graph
Retrieval starts with deterministic lexical candidates, applies validity, trust, and clearance rules, then follows a one-hop entity connection to surface a related stale-document episode with citations. The trace also shows when policy leaves the packet empty and the agent should abstain.
- The event log closes the loop
The final integrity checkpoint records the event count, hash-chain verification, and memory-record totals. It makes the append-only claim concrete: nothing enters or leaves without a trace.
- What this experiment does not do
Memory OS v0 uses deterministic lexical candidate finding and a one-hop temporal entity index. It does not yet use embeddings, a vector database, background consolidation, or a learned policy.