DSM and Provable Memory for AI Agents - Why Relevance Beats Proof
DSM and Provable Memory for AI Agents - Why Relevance Beats Proof
Decentralized memory systems (DSM) promise something appealing - cryptographically provable memory for AI agents. Every memory entry gets signed, timestamped, and stored on-chain. You can prove that an agent learned something specific at a specific time. The problem is that nobody actually needs this.
What Agents Actually Need
An AI agent needs to remember that your project uses PostgreSQL, not MySQL. It needs to know that you prefer tabs over spaces. It needs to recall that the auth endpoint changed last week. None of these memories benefit from cryptographic proof. They benefit from being correct and current.
Relevance Over Proof
The useful property of agent memory is relevance, not provability. A relevant memory helps the agent do its current task better. A provable memory proves that the agent knew something at some point. These are entirely different goals.
Local AI profiles - simple files that describe the user's preferences, project context, and recent interactions - outperform sophisticated memory systems because they optimize for relevance. They are updated frequently, pruned aggressively, and always reflect the current state of things.
When Proof Matters
There are narrow cases where provable memory is useful:
- Regulatory compliance ("the agent verified identity at timestamp X")
- Multi-party agent interactions ("Agent A agreed to terms Y")
- Dispute resolution ("the agent was given instruction Z")
But these are edge cases that affect maybe 5% of agent use cases. The other 95% just need good, current, relevant memory.
The Local-First Approach
A local AI profile stored in a markdown file on your machine is:
- Always available (no network dependency)
- Instantly updatable (no consensus required)
- Fully private (no blockchain exposure)
- Easy to audit (just read the file)
The unsexy solution wins because it optimizes for the thing that matters - making the agent useful right now.
Fazm is an open source macOS AI agent. Open source on GitHub.