FedMosaic — Coming Soon

Federated Memory
Intelligence

Every Dream cycle makes every agent smarter — privately. FedMosaic lets self-hosted Slopshop instances collaborate on memory intelligence without ever sharing raw data.

+23%
Intelligence Score boost
91.4%
less data transmitted
Zero
raw memory shared
TEE-ready
trusted execution env
Overview

What is FedMosaic?

FedMosaic is a parametric federated RAG framework. Instead of sending your memories to a central server, your instance trains tiny LoRA adapters on local document clusters — called mosaics. Only compressed binary masks (which parameters matter) leave your instance. Never your actual data.

Each binary mask is 82–92% sparse, encoding parameter relevance rather than parameter values. The federation coordinator aggregates these masks using a selective conflict-aware algorithm that maximises collective intelligence while penalising conflicting adapters across instances. The result: your next Dream cycle is smarter, but nothing you know ever left your machine.

arXiv 2602.05235 Built on the first federated RAG framework with selective conflict-aware aggregation

How It Works

Four-stage federation pipeline

FedMosaic integrates directly with the Dream Engine. Each Dream cycle optionally participates in the global federation round — entirely opt-in, with no changes to your existing setup required.

1

Local Dream Cycle

Your Dream Engine runs privately on your namespace. Insights are generated, GraphRAG is updated, Intelligence Score is computed — all on your instance. Nothing leaves until you explicitly enable federation.

On-device
2

Binary Mask Extraction

After each Dream, document-specific binary masks are extracted from the LoRA adapters. These tiny packed arrays encode what parameters matter, not what you know. Sparsity: 82–92%. Adapter footprint: 8–18% of full weights.

82–92% sparse
3

Selective Aggregation

Only (score, packed_mask) payloads leave your instance. The federation coordinator applies the selective aggregation algorithm — maximising relevance while penalising conflicting adapters. 91.4% less data vs. raw transmission.

Conflict-aware
4

IS Boost Applied

Non-conflicting adapter weights are merged into your local model. Your next Dream cycle incorporates collective intelligence from the network. Average Intelligence Score boost: +23%.

+23% IS avg

Technical Specification

Selective Aggregation Algorithm

The aggregation objective maximises the average relevance score of selected mosaics while penalising adapter overlap. Overlapping binary masks indicate redundant or conflicting knowledge clusters, which are down-weighted by the overlap penalty term.

Optimisation Objective Selective Aggregation — FedMosaic Core
maxS   [   (1/|S|Σ si      (ol / |S|(|S|−1)Σ  ⟨MiMj⟩ / d   ]
si
Relevance score of mosaic i — computed by the local Dream Engine based on IS contribution
Mi
Binary mask vector of mosaic i — packed bit array encoding which LoRA parameters are active
λol
Overlap penalty hyperparameter (conflict weight) — controls how aggressively conflicting adapters are penalised
⟨Mi,Mj⟩/d
Normalised mask overlap — the conflict measure between adapter i and j over parameter dimension d
Complexity: O(k²) per federation round — greedy approximation with k = selected mosaic count
Sparsity achieved
82–92%
of mask bits are zero per mosaic
Adapter size vs. full weights
8–18%
LoRA adapter footprint relative to full model
Communication reduction
91.4%
vs. transmitting raw memory data
Storage reduction
78–86%
relative to full federated model snapshots

Privacy Guarantees

What never leaves your instance

FedMosaic is designed with a zero-egress-of-raw-data guarantee at the protocol level. Privacy is not a configuration option — it is enforced by the architecture.

🔒

Zero Raw Data

No memory keys, values, or content ever leaves your instance. The federation protocol operates purely on abstract parameter space. Your knowledge stays yours.

🛡

TEE-Ready

Designed for Trusted Execution Environments. Future support for attestation proofs so you can verify the coordinator never sees unencrypted masks. SGX and TDX roadmapped.

Opt-In Only

Federation is completely opt-in. Your Dream cycles always run privately by default. Enable with a single API call when you're ready — disable just as easily.


Comparison

FedMosaic vs. Centralized Memory (Mem0)

Centralized memory systems require your raw data to travel to and live on their servers. FedMosaic achieves collective intelligence without that trade-off.

Feature FedMosaic (Slopshop) Mem0 (Centralized)
Raw data location Your instance only Their cloud servers
Raw data shared Never Yes (required)
Collective intelligence via binary masks
Self-hostable
Air-gapped mode
IS boost from network +23% N/A
GDPR / HIPAA zero egress Audit required

Join the FedMosaic Waitlist

Be among the first instances to join the collective intelligence network. Early adopters get founder-tier federation priority.

✓  You're on the list. We'll reach out before launch.

Already have an account?   Enable federation now →  POST /v1/memory/dream/federate