YantrikDB: a memory database that forgets, consolidates, and detects contradiction

YantrikDB bills itself as a cognitive memory engine for AI agents — not just another vector store. It thinks about what it stores: memories decay over time, similar notes collapse into canonical memories, and contradictions are flagged and surfaced. It has been reported that the project is available as a single Rust binary, a Docker/Kubernetes-ready network server, an MCP plugin for agent frameworks, and as embedded libraries for Python and Rust.

What it does

The pitch is simple and neat. Forgetting is intentional: memories get a configurable half-life so low-importance facts quietly fade. Consolidation collapses redundant fragments into a smaller set of canonical memories. Contradiction detection surfaces factual conflicts (“CEO is Alice” vs “CEO is Bob”) so agents can stop arguing with themselves. There’s also an entity graph, session-aware context surfacing, personality derivation from memory patterns, and a multi-signal scoring formula (recency × importance × similarity × graph proximity). Clever? Yes. Practical? Very possibly.

Deployment and performance

You can run it locally via Docker, hook it into MCP-compatible agents (Claude Code, Cursor, Windsurf), or embed it in apps. Security and ops features are baked in: AES-256-GCM at-rest encryption, Prometheus metrics, per-tenant quotas, HA cluster configs, and runtime deadlock detection. It has been reported that on a modest 2-core LXC cluster with ~1.7k memories recall p50 is ~112 ms (p99 ~190 ms), batch writes ~76/sec, and pre-computed-embedding recalls drop to ~5 ms. The repo also claims a recent 42-task hardening sprint, chaos-tested failover and a hardened alpha v0.5.11; these are reported by the project.

Why it matters

The problem YantrikDB attacks is familiar: dump everything into a vector index and hope retrieval helps. That approach is search, not memory. Real memory is selective, compressive, and self-updating. Could deliberate forgetting and consolidation be the antidote to noisy agents and creeping hallucinations? Maybe. The idea feels like a little spring cleaning for agent context — Marie Kondo for embeddings. Caveat emptor: it’s still early-stage software, and the claims come from the project itself. But if your agent is drowning in context, this is one to watch.

Sources: github.com/yantrikos, Hacker News