KafGraph · Distributed Agent Memory

kafGRAPH

The shared brain for collaborating agents.

A self-owned, persistent knowledge graph in Go. Ingests every conversation and decision in your agent group, exposes it back through brain_search, brain_recall, brain_capture. No agent starts from zero.

Shared Memory. Built for teams. 100% Open Source.

One brain. Many agents.

Distributed memory graph for AI agents. Self-owned, persistent, queryable. Every conversation makes the team smarter.

Most agent memory systems are private and per-user. KafGraph is the first that's shared, distributed, and self-owned.

Why this matters

Multi-agent systems fail at memory before they fail at anything else.

The category-leading agent memory systems (Mem0, Zep, Letta, Cognee) are built for a single user or a single agent. Once you scale to a team of collaborating agents (researcher, coder, reviewer, orchestrator), the memory layer becomes the bottleneck. The research is unambiguous on this.

36.9%

of multi-agent failures stem from inter-agent misalignment, not from model errors. Different agents operate on different versions of reality, leading to conflicting decisions and silent contradictions. A shared memory layer turns those mismatches into resolvable graph queries instead of latent bugs.

Source: Cemri et al., via MongoDB analysis

40–80%

observed failure rates across popular multi-agent frameworks in production traces. Work duplication and stale context dominate the failure modes, with downstream agents acting on assumptions their upstream peers have already invalidated. Persistent shared memory eliminates both classes.

Source: 200+ execution traces, MongoDB engineering

#1 gap

named the most pressing open challenge in agent infrastructure. A recent computer-architecture-style survey identifies cache coherence across agents and structured memory access control as the protocol-level gaps no existing memory system addresses. Single-agent memory engines cannot resolve them by definition.

Source: Multi-Agent Memory survey, arXiv 2603.10062

The brain tool API

Seven tool calls. Any LLM agent runtime.

Agents talk to KafGraph through standard JSON-schema tool calls. Two transports: direct HTTP for embedded mode, or KafClaw skill routing for distributed mode (auto-discovered via roster). The tool surface is deliberately small. Seven primitives that compose into anything an agent team needs.

Tool
Purpose
When agents call it
brain_search
Semantic search across the knowledge graph. Vector similarity over Message, SharedMemory, LearningSignal, Conversation nodes.
"Has anyone evaluated this vendor before?"
brain_recall
Load accumulated context at session start. Edge traversal across conversations, decisions, feedback, threads.
No more zero-start. Every session inherits team history.
brain_capture
Write insights into the brain. Creates LearningSignal nodes auto-linked to the agent and any referenced nodes.
Mid-task: "this approach failed because…"
brain_recent
Time-windowed activity browsing with type and agent filtering.
"What did the team do in the last 12 hours?"
brain_patterns
Recurring themes surfaced via tag aggregation across LearningSignal nodes with occurrence thresholds.
"What concepts keep coming up together?"
brain_reflect
Trigger reflection on demand. Creates a ReflectionCycle, gathers and scores LearningSignals.
End of sprint, end of incident, end of session.
brain_feedback
Submit human feedback on a reflection cycle. Score overrides flow back into LearningSignal nodes.
A human reviews what the brain learned and confirms or corrects.

How you run it

Embedded for one agent. Distributed for a team.

Same binary. Same brain tools. Same OpenCypher and Bolt v4 on the wire. The deployment posture flips with a config flag.

Mode 01 · Embedded

A local brain for a single agent.

KafGraph runs in-process or as a sidecar. BadgerDB stores the graph on local disk. The Brain Tool API listens on localhost. Useful for per-agent personalization, coding agents on a developer laptop, or air-gapped single-agent deployments.

Storage
BadgerDB (LSM, ACID)
Transport
HTTP localhost
Footprint
Single binary, <30MB RAM idle
External deps
None

Mode 02 · Distributed

A shared brain for an agent team.

All agents in a KafClaw group share a single brain across the cluster, partitioned by agentID via FNV-1a hashing, queryable cross-agent. Gossip-based membership via hashicorp/memberlist; internal RPC fan-out merges results. Tolerates partial node failure. This is the mode no other agent memory system ships.

Membership
Gossip · memberlist
Partitioning
agentID FNV-1a
Routing
Fan-out RPC + merge
Co-location
KafScale broker nodes

How the brain compounds

Three loops. Every conversation makes the team smarter.

A private brain only compounds from one agent's experience. A shared brain compounds from the entire team's experience. KafGraph closes the loop in three places.

Loop 01

Automatic ingestion.

Every KafClaw group conversation, skill request, audit event, and shared-memory write becomes a graph node with embeddings. Reads flow directly from S3 segments via the KafScale processor. No broker round-trip, no replay storms.

Loop 02

Reflection cycles.

Daily, weekly, and monthly reflection runs score recent activity by impact, relevance, and value contribution. Weekly and monthly cycles roll up the previous tier. The brain learns what mattered, not just what happened.

Loop 03

Human feedback.

Reflection cycles transition into a feedback state machine (PENDING → NEEDS_FEEDBACK → REQUESTED → RECEIVED). A human confirms what had real positive or negative impact. Score overrides flow back into the LearningSignal nodes.

The four-layer agent stack

KafGraph is the brain. The other three layers carry, run, and watch it.

Scalytics ships four open-source layers that compose into a complete production agent platform. Each is independently useful. KafGraph is the memory substrate. The place where collective experience becomes searchable, reflectable knowledge.

Layer · Transport

KafScale

S3-native, Kafka-compatible streaming. Stateless brokers, permanent memory, infinite retention at object-storage cost.

Layer · Runtime

KafClaw

Agent groups, skill routing, policy mesh, tool registry, shared memory topics. The conversation layer.

Layer · Brain

KafGraph

The distributed shared brain. Property graph, reflection engine, and Brain Tool API. Where collective experience becomes knowledge.

Layer · Observability

KafSIEM

Streaming threat detection and audit on top of the same Kafka backbone. Every agent action carries a citation.

→ Each layer is Apache 2.0. Use one. Use all four.

Honest comparison

Where KafGraph fits among the agent memory systems people actually evaluate.

All five tools below are good at what they're built for. The table is meant to help you choose, not to attack. Self-row highlighted so you can see exactly where the trade-offs land.

System Self-host Graph-native Multi-agent shared External deps License
Mem0 Yes (Apache 2.0) Optional (Pro tier paywall on managed) No · per-user Vector DB + LLM Apache 2.0
Zep / Graphiti Graphiti only · CE deprecated 2025 Yes · temporal KG No · per-user Neo4j or FalkorDB or Kuzu Apache 2.0 / cloud
Cognee Yes Yes · graph-first No · per-user Neo4j or FalkorDB or Kuzu Apache 2.0
Letta (MemGPT) Yes No · tiered memory No · per-agent PostgreSQL Apache 2.0
LangMem LangGraph only No · key-value No · per-thread LangGraph runtime MIT
Claude memory · ChatGPT memory No · vendor-only No No · per-user, siloed Vendor API Closed
KafGraph Yes · single binary Yes · OpenCypher · Bolt v4 Yes · gossip cluster None · embedded BadgerDB Apache 2.0

Pick what fits. Mem0 is the broadest ecosystem. Graphiti has the strongest temporal model. Cognee is graph-first. Letta is tuned for long-running single agents. LangMem is the path of least resistance inside LangGraph. KafGraph is the only one designed from day one for teams of agents that share a brain, with no external graph database to operate.

Build status

Open beta. Feature-complete. Hardening in progress.

Open Beta

All nine layers run. The last 10% is enterprise hardening.

Foundation, server, ingest processor, query engine, brain tools, reflection, feedback loop, distributed cluster, and the embedded management UI are all implemented and exercised by end-to-end tests. The remaining work is enterprise hardening: TLS posture, encryption at rest, OpenTelemetry coverage, Helm packaging polish, and load-testing at scale. We do not call this GA until that work lands. We do recommend it for evaluations, design partners, and non-production agent deployments today.

Apache 2.0 · self-owned Go 1.25 · static binary 602+ tests · race detector on CI Status snapshot: 2026-Q2

Red lines

What KafGraph is not.

We're explicit about the cases we don't serve. If you need any of these, reach for a different tool, and we'll tell you which one.

Not

A general-purpose graph database.

KafGraph implements a deliberate OpenCypher subset focused on agent memory access patterns. If you need full Cypher with stored procedures and a query planner that competes with Neo4j Enterprise, use Neo4j or Memgraph.

Not

A pure vector database.

Vector search exists inside KafGraph because agents need it for semantic recall, not because we're competing with Pinecone or Weaviate. If your workload is "millions of embeddings, no graph," use a real vector DB.

Not

A drop-in Neo4j replacement.

We speak Bolt v4 and a Cypher subset so existing Bolt clients and dashboards work for inspection. We are not a 1-to-1 swap for an existing Neo4j deployment with custom procedures, plugins, or APOC dependencies.

Not

GA software, today.

Open beta means feature-complete on the core scope but hardening still in flight. Use it for evaluations, design partnerships, and non-production agent work. Talk to us before you commit a regulated production workload to it.

Who builds KafGraph

The team that built the federated execution framework now an Apache Top-Level Project.

Scalytics was founded by the original creators of Apache Wayang, the federated data-processing engine. The same engineers maintain KafScale, the S3-native Kafka-compatible streaming platform, and KafClaw, the agent runtime that consumes from it. KafGraph is the third layer in that stack: the brain. The same release discipline (race detection, coverage gates, security scanning) that ships Wayang code through the Apache process applies here.

Scalytics consulting work spans regulated industries: defense, finance, healthcare, energy. We've designed and operated production streaming and AI infrastructure for years before agent runtimes existed. KafGraph exists because our customers needed a memory layer that survives the same compliance review their Kafka does. Read the full team and history.

Apache Wayang PMC KafScale maintainers KafClaw maintainers Apache 2.0 only Confluent partner

Give your agents a brain that compounds.

Read the design notes, run the binary, or talk to the team that's built it. KafGraph is open beta, Apache 2.0, and your data never leaves your infrastructure.