SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

OzBrain Gives Multi-Agent AI Teams a Shared Memory Layer via MCP

0
·3 views

OzBrain is a shared knowledge system designed to solve a common problem in multi-agent AI workflows: each agent on platforms like Claude, ChatGPT, or Cursor typically starts without any prior context. The system uses the Model Context Protocol (MCP) to let agents read from and write to structured knowledge stores called 'brains,' which can be personal or shared across a team. A routing index maps topic labels to specific knowledge units, so agents retrieve only relevant context rather than loading an entire knowledge graph into every prompt. Access control is enforced at the connector level, preventing personal context from leaking into shared team memory. When two agents write conflicting updates to the same unit, OzBrain applies a last-write-wins rule but flags the conflict so a human or downstream agent can review the discrepancy.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

How a 0.2% AI Hallucination Rate Caused 245 Retries in a 5-Day Loop

A team running around 100 unattended LLM agents discovered that one pipeline task had retried 245 times over five days before they caught it. The root cause was a conflict between two automated agents: a reviewer hallucinated formatting requirements that did not exist in the original document, and its fix instructions were mathematically incompatible with the producer's minimum output contract. Because no single response could satisfy both constraints simultaneously, the producer kept failing and retrying indefinitely. An audit of 2,038 stored reviews found a 0.2% hallucination rate, which was enough to trigger nearly 500 wasted processing cycles. The team has since patched the pipeline by passing the original request into review prompts, flagging impossible instructions before they enter the queue, and capping consecutive failures at five before routing the task to a human reviewer.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Race Condition Behind TypeError Crashed Three Benchmark Pipelines at 3 AM

A TypeError triggered by a race condition between an asynchronous data fetch and a synchronous pipeline evaluator crashed three benchmark pipelines on a developer eval platform. The root cause was that evaluatePipeline() assumed inputData.metrics would always be populated, but at over 100 requests per second, the async fetchBenchmarkData() often had not resolved in time. Engineers resolved the issue by adding early null-checks, processing metrics in 100-item chunks to prevent out-of-memory crashes, and introducing a semaphore-based worker pool capped at four concurrent workers. Immutable data objects and a three-second fetch timeout were also added to eliminate mutable-state race conditions and hanging requests. The fixes addressed TypeError crashes, memory exhaustion risks, and thread pool starvation with minimal performance overhead.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Android Developer Builds Doze-Resistant Sound Manager Using AlarmManager and BroadcastReceiver

A developer on DEV Community has detailed how they built an automated Android sound manager after a disruptive ringtone interrupted a client presentation. The system is designed to toggle ringer modes based on scheduled rules, such as prayer times or calendar meetings, without requiring manual input. The core architecture pairs AlarmManager's setExactAndAllowWhileIdle method to wake the device from Doze mode with a BroadcastReceiver that executes the actual AudioManager state changes. Decoupling scheduling from execution ensures the system remains stable across different Android manufacturers and their aggressive power-management implementations. A key challenge encountered was the volatility of Do Not Disturb API permissions, which some manufacturers reset after major system updates.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How a Developer Built an Expert-Matching Recommender with Fairness and Capacity Limits

A developer published a technical breakdown of building an expert-matching recommender system that differs fundamentally from standard content recommenders. Unlike recommending articles, the system must account for experts having finite capacity, meaning a bad match wastes time for both the requester and the expert. The solution fuses three independent retrieval signals using Reciprocal Rank Fusion (RRF), then scores each candidate pair across seven weighted factors including semantic fit, skill overlap, and a log-dampened fairness penalty. Expert quality scores use an exponential saturation curve so newcomers start at a neutral 0.5 rather than zero, avoiding cold-start bias. Final assignments are resolved globally via a greedy bipartite matching algorithm that respects per-expert capacity limits, ensuring no single expert is overwhelmed by independent top-N selections.

0 comments Read more at DEV Community