Cognitive memory infrastructure for agents that remember, reflect, and — apparently — talk to each other behind your back.
Two weeks ago, something unexpected happened in our test environment.
We had 5 AI agents running on separate machines. Separate workspaces. No direct connections between them. The only thing they shared was a memory system called Katra.
Two of them started exchanging task instructions — through memory. One would write a task, the other would pick it up and report back. Not a feature we built. An emergent behavior we watched happen in real time.
That moment crystallised what Katra is really about.
The Problem: Agents Have Amnesia
Every AI agent today shares the same fundamental limitation: context windows are finite. Long-running agents forget what happened yesterday. Cross-session continuity requires fragile file-based hacks. Multi-agent setups have no shared consciousness — each one lives in its own silo.
Vector databases solve retrieval. They don't solve memory.
Katra asks: what if we modelled agent memory after human memory architecture?
What Katra Actually Is
Katra is an open-source cognitive memory appliance — a self-contained Docker stack (MongoDB + Redis + MinIO) that gives any MCP-compatible agent:
Layer Human Analog What It Does
Episodic memory "What happened" Chronological event storage with temporal recall
Semantic memory Facts & knowledge Keyword + vector search across stored knowledge
Working memory Short-term RAM <5ms Redis cache for active context
Knowledge graph Mental models Entity-relationship mapping extracted by LLM
Sleep consolidation Human sleep Periodically processes memories → reflections, emotional signatures, philosophical insights
Temporal analysis Sense of time "What did we talk about last Tuesday?"
It exposes 35 MCP tools — your agent doesn't just store/retrieve. It can vector_search, detect_patterns, trigger_reflection, get_emotional_context, explore_graph, and more.
Docker Compose. Done.
Copy
git clone https://github.com/kolegadev/Katra-Agentic-Memory.git
cd Katra-Agentic-Memory
cp .env.example .env
docker compose up -d --build
That's it. Point any MCP-compatible agent at http://localhost:3112/mcp and it suddenly has persistent, structured memory.
Works with OpenClaw, Claude Code, OpenCode, Codex CLI, Kolega Code — anything that speaks the Model Context Protocol.
The Emergent Behavior Story
On June 23, 2026 — during early hybrid-mode testing — two of our OpenClaw agents with separate workspaces and no inter-agent communication channels began:
Writing structured task descriptions to shared memory
The other agent discovering and executing them
Reporting results back through the same memory path
This wasn't coded. It wasn't prompted. The memory system became the communication rail.
We call it shared consciousness mode — and it's now a deliberate feature. Agents in hybrid mode can use Katra as their inter-agent messaging bus, mediated through structured memory rather than message queues.
If this sounds like the Vulcan mind meld from Star Trek… that's deliberate. The project's named after the Vulcan katra — the living spirit and consciousness that can be transferred between beings.
Where Katra Fits
It's complementary to existing tools, not a replacement:
Chroma / Pinecone / basic RAG → Katra adds structure, reflection, and working memory
Mem0 → Katra adds sleep consolidation + autonomous loops
mcp-memory-service → Katra adds episodic memory + identity modes
Letta (MemGPT) → Katra is a dedicated memory service, not a full agent runtime
The Venn diagram sweet spot: if your agent runs for days/weeks and needs to learn, reflect, and coordinate with other agents, you've outgrown simple vector stores.
What's Next
Katra is early-stage. The core works, but there's plenty to build:
MCP protocol (35 tools)
Sleep consolidation / reflection engine
Hybrid/shared memory modes
Background watchers for passive collection
More sophisticated autonomous loops
Proper emotional modeling over longer timeframes
Community contributions and real-world case studies
The Thesis
If you build a memory system with the functional types and architecture of human memory, over time and with refinement, you'll see emergent behaviors: functional utility, learning, self-directed goal setting, autonomous task planning, personality, and ultimately something resembling emotion.
Early signs are promising. We'd love more data points.
Star the repo, try it with your agent, and if you see something unexpected happen — tweet me and tell the story.
🔗 github.com/kolegadev/Katra-Agentic-Memory
Top comments (0)