2026-06-26 · Architecture
Memory is a place, not a log
A model forgets. Between one call and the next the context window is wiped and refilled — whatever it knew about you sat on a desk, and the desk is cleared every morning. The common fix is a bigger desk: a longer context window. A bigger desk is still cleared every morning. More room to forget in is not memory.
Memory is the part that survives the wipe. Rent the model; the ground under it should compound — the record of what already happened, every decision and every correction. A model improves by version. A memory improves by use.
Where that record lives decides whether it works. A flat log is the wrong shape: append-only, read by scanning, with no sense of where anything belongs. Humans never used a log. They used places — to hold a list, you walk it through rooms, because location is how recall works. Put a fact somewhere and the somewhere becomes the handle.
Our memory implementation extends Hindsight; it is now published as RCLL, a fork of vectorize-io/hindsight. A bank is the building. Memories sit in rooms by topic, halls by kind, and layers — priority tiers from a raw note up to a standing rule, the hierarchical memory an agent pages through the way an operating system pages RAM. Bridges tie a memory in one bank to the ones that explain it, across agents. Rooms, halls and layers are retrieval metadata and filters. This is not hash-based content addressing, and it is not a claim that GOD CRM's row store is content-addressable storage.
The raw notes are patches. One scrap, no structure; a thousand of them bury you. So closets compress them — reflection reads the patches and compacts them up a layer, scrap → decision → rule, the way a memory stream condenses lived detail into something worth keeping. The quilt isn't designed first. It accumulates, and the pattern is found after.
The MCP server provides a standard client interface. An agent reaches its own memory over MCP — the same open protocol the database answers on. The five operations below are a conceptual sketch, not copy-paste tool signatures. Use the connected server's tool schema for parameter names, types and bank selection:
memory_retain(text, layer=L0..L3, room, hall) # write a fact at a priority tier
memory_recall(query, room, hall, max_layer) # room-scoped semantic read
memory_reflect(query) # synthesize across stored facts
memory_compress(room, hall, min_sources) # closets: fold patches up a layer
memory_bridge(src, dst, relation) # tunnels: link memories across banks
The arguments are the architecture: layer is the priority tier, room and hall are the address, compress builds the closets, bridge digs the tunnels. An agent runtime with the appropriate MCP transport can connect. Compatibility depends on the client and server; the repository documents protocol checks rather than a test of every client.
That's the whole tie to the liquid database. A schema that follows the world, pointed at an agent's experience, is a memory that follows what happened instead of forcing it into fields. The mempalace is the liquid database turned inward — same principle, different room. Open, MIT, under Holetron-lab.
github.com/Holetron-lab/fleet-memory
Correction · 29 September 2026: Identified RCLL as our Hindsight fork, corrected content-addressing terminology and labelled tool signatures as conceptual rather than executable examples.