1. CoBrA: A Self-Evolving 24/7 Agentic Platform with Lossless Memory and Infinite Context

CoBrA (CoMeT Browsing Agent) is a multi-agent orchestration platform built natively on CoMeT (Cognitive Memory Tree), our open-source memory protocol. It's designed to solve the fundamental scaling wall every long-running AI agent hits today.

2. The Problem: The Agent Memory Crisis

Anyone building agents that run beyond a few dozen turns has hit this:

Tool Bloat: Raw tool I/O crammed into context creates a garbage-in, garbage-out loop. By 300+ turns a session's raw prompt crosses 54M tokens — 54× past any frontier model's 1M window, forcing aggressive truncation or reasoning collapse long before then.

• Session Silos: Memory dies when the session ends. The agent re-learns everything from scratch.

• The Token Tax: Every tool log and search result dumped into the prompt spikes cost and kills reasoning — "Agent Dementia," where the model forgets its mission while staring at logs.

• The Similarity Trap: Flat RAG retrieves what's mathematically similar, not what's logically relevant. No narrative structure, no task awareness.

The Solution: CoMeT: The Memory Philosophy

The core idea is simple: stop treating the context window as memory.

We replace bloated context windows with a compacted, hybrid (index + graph) memory database, and we replace the memory paradigm itself with fast, agentic tool calling. The active context stays fresh — holding only the last 2-3 turns plus compacted memory references (id, summary, trigger). Memory is recalled by tool, not by stuffing.

When a task requires historical context or detail, the agent dynamically recalls only the necessary resolution of memory: summary for orientation, detailed for reasoning, raw for precision. Five layers, structured by a lightweight Compactor SLM:

• Summary — high-level overview

• Detailed Summary — deeper context for complex reasoning

• Trigger — conditions that activate the memory

• Tags — metadata for rapid retrieval

• Raw Source — untouched original data

A separate Sensor SLM monitors sessions, browsing, and tool calls in real time, deciding when an interaction is worth persisting in the first place.

It's fully compatible with current agent systems — but as a drastically more personalized and focused memory engine.

3. CoBrA: The Platform

CoBrA is what happens when you build an agent runtime natively on CoMeT:

Infinite session length — single sessions run indefinitely with lossless compact/recall, handling hundreds of millions of characters of context.

• Project-based multi-agent orchestration — each agent maintains its own CoMeT memory tree while sharing nodes, talking to each other across the session.

• 24/7 unattended operation — goal-driven convergence for long-run research workflows and optimization task (targeting metrics like Chamfer distance, volumetric IoU).

• Early self-evolution — agents improving their own workflows across sessions without human intervention, based on memory evolution.

• Memory-native harness — prompts co-designed with CoMeT's graph (tag-based pre-flight for prior failures, auto-regenerating session briefs on user corrections), so the agent reasons through memory instead of around it.

4. Production Use Cases

• CAD AI Harness Engineering — 2D→3D parametric conversion, drawing-to-code, assembly generation. #1 on our internal benchmarks globally.

• End-to-End Robot Agent — text-to-CAD pipelines for complex assemblies (robot arms, turbine blades, humanoids) with part search, CAD memory, and automated assembly.

• Lesson based self-improvement: agent take action very quickly, store all those experiences.

5. Roadmap

• Memory Market — a marketplace where users share Memory Maps, letting other agents leverage accumulated workflows and domain expertise.

• Memory-Augmented Reasoning — using the memory graph as a reasoning substrate, not just a retrieval store.

CoMeT (open source): https://github.com/Dirac-Robot/CoMeT

Built by The Dimension Company — the CAD AI platform for defense and automotive.