Neruva is the custom memory & grounding substrate I run across everything I build — Cairn, SimGen, FavourBee, and my day-to-day operations — through Claude Code MCP. A records store, a knowledge graph, and deterministic snapshot/replay: reproducible, auditable recall where the same query gives the same answer and every decision can be replayed. Built for my own work first and proven there daily — open to serious pilot and licensing conversations.
Runs on Claude Code MCP · ~25 tools over api.neruva.io · live demo on request
# how my agents remember — one substrate, every project $ claude mcp add neruva # remember once, recall deterministically — anywhere agent_remember("Kai's tenant lease ends 2026-03-31") agent_recall("when does Kai's lease end?") → 2026-03-31 · cited · same answer every run
The split is deliberate: the server is a deterministic substrate — storage, retrieval math, provenance, counts — while meaning and judgment stay with your agent. That boundary is what makes recall reproducible, corrections enforceable, and every decision replayable. The same trait that makes a cryptographic receipt verifiable, applied to memory.
Every serious agent deployment gets asked this eventually — by a regulator, an auditor, or a customer. Logs show what an agent output. Almost none can show what it knew when it decided, or replay the decision faithfully. With EU AI Act high-risk enforcement live and MCP still lacking a standard audit trail, that gap is now a compliance problem, not a curiosity.
Each agent action signs a record committing to the SHA-256 of the exact context it read.
The context lives in Neruva, content-addressed — the address is the hash, verified server-side on write.
An auditor, knowing only the committed hash, fetches the exact bytes back. A corrupted store read fails closed.
The decision re-runs against the restored context and must reproduce the committed output byte-for-byte.
Proven on the live API with two independent agents: one agent's decision was reproduced and audited by a second agent that held nothing but the hash — byte-for-byte, no shared state — and an agent-driven tamper failed closed. Authority (who authorized the agent) is verified by the offline verifier from Cairn. Named limit: byte-for-byte replay covers deterministic decisions; LLM actions require committing model, parameters, and prompt as context.
Everything an agent needs to remember well — exposed as MCP tools over api.neruva.io. The newest is replayable agent audit.
Append-only typed events with semantic + BM25-RRF recall. Ingest, query, timeline, compact, export to a portable .neruva file. The substrate auto-embeds text server-side.
agent_remember / agent_recall / agent_context — federated retrieval across records and the knowledge graph, with cross-session fan-out and a paste-ready context block.
Subject–relation–object triples with temporal validity. Exact multi-hop neighbors, reverse lookups ("who controls X?"), and corrections via replace-fact that keep the prior state as history.
Snapshot a namespace to an immutable blob and restore it bit-for-bit from a seed. Time-travel queries against historical state — the same inputs always reproduce the same answer.
Tell it a fact is wrong once and the correction is enforced — recalled before extraction and injected as a mandatory override. Not retrained. It does not recur.
GDPR / Quebec Law 25 deletion: forget records by kind/tag/time/user, or hard-delete every fact about an entity in both directions — cleanly, with the audit trail intact.
Content-addressed context store: an agent action commits to the hash of what it read, and any flagged decision is reproduced byte-for-byte from the exact memory it was made with — the newest layer, detailed above.
My projects and operations share a single memory layer — so decisions, corrections, and history carry across all of them, and I can replay any of it.
Neruva is the memory substrate behind my work — a working prototype proven in daily production. I'll demo it live over MCP: records, the graph, deterministic replay. The rest of my work is on my portfolio.