RCLL

Self-hosted shared memory for a team of AI agents. Storage + structure in one system.

RCLL — team memory for agent fleets. Built on Hindsight (github.com/vectorize-io/hindsight, MIT).

[!IMPORTANT] You are looking at a read-only mirror. The canonical repository is godcrm.ai/git/holetron-lab/fleet-memory — self-hosted, public, clonable anonymously with no account anywhere in the chain. Everything here is pushed out from there, so a merge performed on GitHub is overwritten by the next sync, usually within the hour. Issues and stars belong here and are read. A pull request is welcome here as well — it gets merged on the canonical side and arrives back here on the next sync.

RCLL is a fork of vectorize-io/hindsight (MIT). It keeps Hindsight's storage engine and adds rooms — topic scoping over one shared store, which is selectivity rather than isolation — plus a hierarchical depth model (L0–L3). The room/hall/layer taxonomy is prior art in the hierarchical-memory space; the implementation here is our own.

RCLL is recall with the vowels dropped — the one operation every agent in the fleet performs before it does anything else. The tool is literally called memory_recall; the product is named after the call.

Its one structural property worth remembering: the read path never invokes a language model. A recall costs CPU and zero model tokens — see architecture.

Status

The source is public and MIT. There is no packaged release yet: fleet-memory-mcp is not published on npm and no container image is pushed. Running RCLL today means building from this tree, which the quick start below does. Don't quote an install command as working until rcll.ai shows one.

How it works

┌──────────────────────────────────────────────────────┐
│                         RCLL                         │
│                                                      │
│  ┌─── Room: auth ───┐  ┌─── Room: pipeline ──┐       │
│  │ Hall: facts      │  │ Hall: decisions     │       │
│  │ Hall: procedures │  │ Hall: events        │       │
│  │ Hall: warnings   │  │ Hall: facts         │       │
│  │                  │  │                     │       │
│  │  L0 ████ always  │  │  L0 ████ always     │       │
│  │  L1 ███░ warm    │  │  L1 ███░ warm       │       │
│  │  L2 ██░░ cold    │  │  L2 ██░░ cold       │       │
│  │  L3 █░░░ archive │  │  L3 █░░░ archive    │       │
│  └──────────────────┘  └─────────────────────┘       │
│           │                      │                   │
│           └──── Tunnel ──────────┘                   │
│                (cross-bank bridge)                   │
│                                                      │
│  Closets: compressed summaries + source pointers     │
└──────────────────────┬───────────────────────────────┘
                       │
              Hindsight vector store
              (embeddings + semantic search)

Rooms — topic isolation. Auth, pipeline, infrastructure, schema — each topic in its own room. An agent searching for auth facts won't wade through 500 deploy memories.

Halls — knowledge typing within a room. Fact, event, decision, procedure, warning. The system knows what it's looking at before reading — like Content-Type for memory.

Layers L0–L3 — four priority tiers. L0 (core) is always loaded. L3 (archive) is deep-search only. Same idea as CPU cache hierarchy: L1 is fast and small, RAM is slow but holds everything.

Closets — AI-compressed summaries with source pointers. Deduplication at the knowledge level: 10 related facts → 1 paragraph + references.

Tunnels — cross-bank bridges between agents. Agent A discovers an insight — Agent B sees it through a tunnel without data duplication.

What this fork adds

This is a list of our additions relative to our branch point (d054b884, April 2026) — not a claim about what upstream Hindsight does today. Upstream has shipped roughly 1,700 commits and four minor releases since we branched; assume anything below has an upstream answer we have not evaluated, and read FORK.md before treating this as a comparison.

Everything is additive: the upstream /retain and /recall contracts as of our branch point still work unchanged, and every new parameter is optional.

Measured

Retrieval quality, our own models on the public LoCoMo dataset, using a third-party harness rather than one we wrote. Full method, the arms that lost, and the caveats: rcll.ai/docs/benchmarks/.

This is retrieval quality, not answer accuracy. It is not comparable to figures of the form "77% on LoCoMo", which measure a reader and a judge on top of a store. We publish no accuracy number because we have not run a reader and a judge.

Two results that go against us are on the benchmarks page rather than left out: on multi-hop questions our default fusion is worse than vector-plus-reranker, and on single-hop BM25 alone beats dense retrieval.

Latency on CPU with no GPU: ~0.29 s for search, ~3.0 s including the cross-encoder reranker. The reranker is 85% of the time and the single largest quality gain we can measure.

Quick start

git clone https://github.com/holetron-lab/fleet-memory.git
cd rcll
cp .env.example .env
# edit .env with your config
docker compose -f docker-compose.rcll.yml up -d

This builds the image from this tree — there is no published image to pull, so the first run compiles and is not fast. The API then listens on http://localhost:5100.

Clients written against upstream Hindsight's API as of our branch point keep working — the added parameters are optional. It is not a drop-in for current upstream Hindsight, which is several releases ahead of this fork.

Embeddings

Ships with BAAI/bge-small-en-v1.5 (384-dim) — fast, CPU-friendly, baked into the image so first run needs no network download. It's English-optimized; recall quality on other languages degrades.

For multilingual memory (e.g. RU, multi-script), point it at a multilingual model:

HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL=BAAI/bge-m3   # 1024-dim, multilingual

Dimension is detected automatically. ⚠️ Switching models changes the vector dimension — do it on an empty memory store, or wipe + re-embed, since existing vectors can't be mixed across dimensions.

Running without an LLM key

LLM_PROVIDER=none is a supported configuration, and it is a smaller product rather than the same one for free: retain drops to chunk mode — chunks stored and embedded whole, with no fact extraction, no entity resolution, no causal links, and consolidation and reflection off. You get a hybrid vector-and-lexical chunk store. Reading is unaffected, because reading never calls a model anyway. Choose this deliberately, or point extraction at a local model — don't arrive here by leaving a field blank.

MCP Server

The mcp-server/ directory contains a standalone MCP server over stdio.

What is actually verified, as of 2026-08-24, against a live backend: protocol version 2025-06-18; initialize, tools/list and tools/call all round-trip; memory_recall returns real results. That is a protocol-level check run directly over stdio — not a client-by-client compatibility matrix.

Any client that speaks MCP over stdio should therefore work, but we have not sat in front of each one. Listed below is the config we run ourselves (Claude Code) and no others. If you get it working with a different client, a PR to this section is the useful kind.

Tools

memory_recall is the only one of the five that never calls a model. memory_reflect is an agentic loop with repeated LLM calls — if you expose this server to anything untrusted, expose memory_recall alone.