Notes on memory you can prove
We're building the trust layer for AI agents — provenance, audit trails, and memory that cleans up after itself. Here's what we're learning as we go: the failure modes, the design calls, and the occasionally uncomfortable truths about how agent memory really behaves in production.
memory decayagent memorybelief dynamicsYour agent's memory is probably mostly garbage
Agent memory doesn't fail loudly. It rots — stale facts outranking fresh ones, contradictions living side by side. Here's why it happens, and what self-cleaning memory does about it.
Harshit··4 min read
provable memoryAI agentstrustWhy we built provable memory for AI agents
Most agent memory asks you to trust a black box. We think memory you can't audit is a liability, not a feature — so we built the one you can prove. Here's the case for provable memory.
Harshit··4 min read
agent memorymulti-tenantprivacyWhose memory is this? The boundary problem in every agent that serves more than one context
The moment an agent talks to more than one customer, user, or project, its memory starts leaking between them. It's a privacy problem and a correctness problem at once — and you can't prompt your way out of it. Here's the boundary that actually works.
Harshit··5 min read
Mem0 alternativeZep alternativecomparisonMem0 vs Zep vs Hakuya: what actually matters in agent memory
A practical comparison of agent memory layers. Mem0, Zep, and vector databases solve storage and retrieval. Hakuya adds the layer they don't: memory you can audit, prove, and erase.
Harshit··4 min read