I built a lightweight memory layer for Claude Code
Claude Code memory without Redis—138 lines of diskcache beats 'use a vector DB' conventional wisdom.

Progressive rehashing kills 200ms pauses, but Redis already does this.
Systems programmers, backend developers building databases
Redis · Memcached · KeyDB
Claude Code memory without Redis—138 lines of diskcache beats 'use a vector DB' conventional wisdom.
TLA+-proven correctness for a Redis-compatible leaderless architecture.
Vector memory for agents beats truncation; Rust core with Python/Node/Go bindings.
Deleted lock-free ring buffers and atomic CAS — database got simpler and better.
Single-file mmap storage plus an HNSW vector index and explicit graph edges is an elegant, practical combo — think "SQLite for agent memory" with CRC-32 crash recovery and zero-server convenience. The C++20 core + nanobind gives zero-copy NumPy views and GIL-free searches, and the claimed FAISS-like throughput makes this genuinely interesting for local setups; main caveat is build/toolchain friction and how rich the surrounding ecosystem becomes.
Jepsen verification and Raft clustering on a Redis wire-compatible Rust database.