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).

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.

One thing worth knowing before you budget context: the backend treats limit on recall as a retrieval hint, not a result cap — it returns everything inside its own token budget, around 110 facts. The MCP layer enforces your limit on the way out, so a limit: 2 recall costs about 1 KB instead of 43 KB. Recall spends no model call; it does spend context.

Setup

cd mcp-server
npm install
FLEET_URL=http://localhost:5100 node server.js

Claude Code config

Add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "rcll": {
      "command": "node",
      "args": ["/path/to/mcp-server/server.js"],
      "env": {
        "FLEET_URL": "http://localhost:5100",
        "FLEET_BANK": "my-agent-bank"
      }
    }
  }
}

Upgrading from the old package name? HINDSIGHT_URL and the legacy bank variable are still read as a fallback, so an existing config keeps working — it just prints a deprecation notice on start.

See mcp-server/README.md for full docs and environment variables.

API changes from upstream

The base /retain and /recall endpoints are backward-compatible with upstream as of our branch point. New parameters are optional.

New parameters

New endpoints

Room/Hall taxonomy

Rooms (topics)

auth · pipeline · infrastructure · deployment · schema · api · ui · tax · hr · legal · compliance · monitoring · agent · general

Halls (knowledge types)

warning · decision · procedure · event · preference · discovery · fact

Layers

Auto-classification

RCLL includes a keyword-based classifier (room_hall_classifier.py) that assigns room and hall automatically when not provided. No LLM call — classification is instant and free.

Extensible: add keywords to ROOM_KEYWORDS / HALL_KEYWORDS dictionaries.

Examples

Store a memory

curl -X POST http://localhost:5100/retain \
  -H "Content-Type: application/json" \
  -d '{
    "bank": "project-alpha",
    "text": "Never restart PROD without confirming staging works first.",
    "room": "deployment",
    "hall": "warning",
    "layer": 0
  }'

Scoped recall

curl -X POST http://localhost:5100/recall \
  -H "Content-Type: application/json" \
  -d '{
    "bank": "project-alpha",
    "query": "deployment safety rules",
    "room": "deployment",
    "hall": "warning",
    "max_layer": 1
  }'

Cross-bank bridge

curl -X POST http://localhost:5100/bridge \
  -H "Content-Type: application/json" \
  -d '{
    "source_bank": "project-alpha",
    "target_bank": "project-beta",
    "room": "infrastructure",
    "hall": "procedure"
  }'

Known limits

Honest list, kept here rather than only on the site:

• You cannot export your memory yet. The upstream export command emits a bank template — config, mental models, directives — and none of your stored content. A full dump that carries rooms, halls, layers and the link graph is the top item on our list, because it is the one thing that should not be missing from a store you self-host.
• No accuracy benchmark. Retrieval quality is measured and published; a reader-and-judge run is not.
• Adversarial / unanswerable questions — where the right answer is "I don't know" — are not covered by the retrieval metric we report, and that is the category most likely to embarrass us.
• Switching MEMORY_MODE later does not migrate what you already stored. Pick before you accumulate.

What we changed

A taxonomy layer over Hindsight's vector store, plus a standalone MCP server.

Key additions:

• room_hall_classifier.py — keyword-based taxonomy engine (new)
• aa1_add_room_hall_to_memory_units.py — DB migration: flat → hierarchical, adds room/hall + layer column (new)
• mcp-server/ — standalone MCP server with 5 tools (new)
• Storage layer — room/hall/layer metadata on every write
• Retrieval — room-scoped search with hall filtering
• Compression — closet generation with source linking
• Tunnels — cross-bank memory sharing protocol

Full architectural spec: RCLL.md. Commit-by-commit account of the fork: FORK.md.

Upstream

This fork branches from d054b884 and its changes are a readable series on top of that commit, so taking a new upstream release is a rebase rather than an excavation:

git remote add upstream https://github.com/vectorize-io/hindsight.git
git fetch upstream --tags
git rebase --onto <upstream-tag> d054b884 main

Expect real conflicts: the rooms work touches the same engine files upstream has rewritten most. FORK.md lists which ones and why.

Credits

• Hindsight by vectorize-io — the memory storage engine
• Holetron — fork maintainers, MCP server, integration

License

MIT — same as upstream Hindsight. See LICENSE.