Hi everyone — I’m an independent researcher working on memory systems for long-lived AI agents, and I’d love technical feedback on a system I’ve built called Ashnode.

Ashnode is a bounded, inspectable, and reproducible memory layer designed to fix a core issue in standard RAG systems: when facts evolve over time, agents can retrieve stale or contradictory context with no clear notion of what’s current.

Key ideas:

• Claim-key supersession: only the latest live fact is returned by default
• Contradiction surfacing using query-grounded NLI
• Completeness flags so agents know when context was capped
• Bounded recall latency under explicit policy caps
• Audit-friendly context packets with provenance and revision metadata

What’s in the work:

• A formal retrieval contract for agent memory
• Mathematical proofs for:
• Bounded synchronous recall latency
• Why exhaustive contradiction detection is infeasible globally and must be query-grounded
• An open-source reference implementation

Some results:

• Warm p95 recall latency: 3.04 ms at 10K items on commodity CPU
• Latency growth only 1.27× for 20× corpus growth
• Synthetic temporal benchmark:
• Plain RAG: 11.3% top-1 temporal accuracy, 78.8% stale fact rate
• Ashnode: 100% temporal accuracy, 0% stale fact rate

Paper + code:

• Repo: https://github.com/itachi-hue/ashnode

If this sounds interesting, comment or DM me — I’d especially appreciate feedback on:

• Novelty / related work gaps
• Whether the evaluation is convincing
• Real-world deployment concerns
• Suitable venues / arXiv category

Thanks!