Skip to content

Blog

Long-form writing about EvolutionDB — what we shipped, why, and how it holds up under the engineering details.

  • :material-newspaper-variant: How we gave Claude a real long-term memory

    2026-05-05 · ~12 min read

    A field note on building EvolutionDB's agent-memory layer for a non-engineering audience: the polling tax, the four-product side stack, the single-binary alternative, and the temporal trick that makes audit-and-replay free.

    Read the product story :material-arrow-right:

  • :material-rocket-launch: Powering long-term memory for agents — v3.0.0

    2026-04-28 · launch announcement

    The v3.0.0 release is the closure of ADR-002: 26 tasks, ~10 weeks, one binary. Native agent-memory DDL, vectors with HNSW, temporal queries, push-not-poll streaming, and six framework adapters.

    Read the launch post :material-arrow-right:

Technical series — Long-term memory in EvolutionDB

A companion to the product story. Eight articles that go underneath the marketing surface and walk through every catalog object, every wire format, every index structure, every fsync.

# Article
00 Series index
01 The MCP bridge — JSON-RPC over stdio in 388 lines
02 MEMORY STORE on disk — DDL, DML, and the catalog tree
03 VECTOR(N) and HNSW — semantic recall against a live transactional workload
04 Temporal memory — FOR SYSTEM_TIME AS OF on top of MVCC
05 Push, not poll — durable subscriptions and CDC streaming
06 The C SDK and FFI — one core, six languages, six framework adapters
07 Multi-tenant memory — namespace hierarchy meets row-level security
08 Benchmarks — what an agent-memory workload actually looks like

A note on tone

The product piece is written for a Medium-style audience and deliberately avoids internal identifiers. The technical series does the opposite — it names files, functions, page types, error codes, and on-disk layouts directly. If you came in looking for a feeling, start with the product piece. If you came in looking for a hex dump, the series is the right place.