// the find
MemoriLabs/Memori
Memori is agent-native memory infrastructure. A LLM-agnostic layer that turns agent execution and conversation into structured, persistent state for production systems. Built for enterprise, Memori works with the data infrastructure you already run, no rip-and-replace, and deploys across managed cloud, single-tenant cloud, VPC, and on-premises.
Memori is a persistent memory layer for LLM agents that intercepts SDK calls (OpenAI, Anthropic, etc.) and automatically extracts structured state — facts, preferences, skills, relationships — without requiring you to change your agent's prompt or logic. The core engine is written in Rust and exposed via Python and Node bindings, with storage backends for SQLite, Postgres, MySQL, and MongoDB. It's aimed at teams building agents that need to remember things across sessions without rolling their own memory pipeline.
The Rust core is the real story here — embedding generation, retrieval, and augmentation pipelines in a compiled binary means the per-turn overhead is low, which matters when you're doing this on every LLM call. The BYODB model is genuinely useful: you point it at your existing Postgres or SQLite and it manages its own schema, so you're not forced into a proprietary data store. The LoCoMo benchmark result (81.95% accuracy at under 5% of full-context token cost) is specific and reproducible — they published the notebooks and the paper. The SDK intercept pattern — wrapping the OpenAI or Anthropic client rather than wrapping prompts — means memory works even in frameworks that don't expose a clean hook point.
The interceptor approach has a fragile dependency on the internal call signatures of each SDK it wraps — every time OpenAI or Anthropic ships a breaking SDK change, Memori breaks silently until they patch it, and they're already maintaining six LLM clients plus streaming variants. The 'Advanced Augmentation' taxonomy (facts, skills, people, relationships, rules, events) sounds thorough but the extraction quality entirely depends on an LLM classifying its own outputs, which means you get whatever hallucinations the extractor model produces stored as ground truth — there's no correction mechanism described. Hermes and OpenClaw are thin wrappers around an undocumented agent runtime; if either of those projects stalls, those integrations become dead weight. Self-hosting on anything beyond SQLite requires standing up the Rust bindings, which adds a native build step to your Python or Node deployment — the docs wave at this without explaining what breaks if your platform doesn't ship a compatible binary.