Developer Builds Shared Memory Layer 'Passport' to Sync Context Across AI Coding Agents

·2 views

A developer built an open-source tool called Passport to solve a persistent frustration: AI coding agents like Claude Code, Cursor, and Codex each start every session with no memory of prior context, forcing users to re-explain the same project details repeatedly. Passport acts as a shared memory layer using an MCP server, allowing one agent to store information that any other agent can later retrieve. A key proof-of-concept moment came when Codex correctly recalled authentication and frontend details that had only been told to Claude Code in a separate session. The project added features beyond basic memory sharing, including provenance tracking to identify which agent contributed which knowledge, conflict detection for contradictory inputs, and user-level memory isolation. Development hit technical hurdles, including a Windows-specific bug where verbose stderr output from the Cognee memory library filled a 64KB pipe buffer and crashed the MCP client during recall operations.

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)

Context Mode Cuts AI Agent Token Bloat by Up to 98% Before It Starts

Context Mode is an open-source, MCP-based context management tool designed to prevent token bloat in AI agent workflows before it occurs, rather than compressing data after the fact. The system intercepts tool output at the source, stripping structural noise from responses such as large DOM snapshots and code search results. In testing, a 315KB Playwright page snapshot was reduced to 5.4KB, a 98% reduction, while a 100-result code search shrank by 92%. The tool also offers session continuity via SQLite FTS5, meaning context persists across restarts, and uses a batch-query approach to reduce redundant read operations. Developers can layer Context Mode alongside other tools like Headroom and tokdiet for compounded token savings across their AI agent stacks.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why small businesses should keep humans in control of AI agent actions

A framework for deploying AI agents in small businesses argues that full autonomy is unnecessary and often risky, especially when agents are connected to real business tools. The model distinguishes between safe read-only actions, draft-and-stage tasks, low-stakes writes, and high-risk irreversible actions such as issuing refunds or publishing content. Tools like n8n enable a human-in-the-loop pattern where an AI agent pauses before executing sensitive actions and routes an approval request via channels like Slack or Telegram. Rather than treating human approval as a sign of incomplete automation, the approach frames it as the feature that makes AI agents trustworthy and deployable for small teams. Sorting agent tools into risk buckets before connecting them to business systems is presented as a practical first step for any business owner.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Solana's Architecture Gives AI Agents a Real Edge Over EVM Chains

Ethereum and most EVM-compatible blockchains impose significant constraints on on-chain AI agents, including roughly 15 transactions per second, 12-second finality, and high gas costs that make autonomous, real-time execution largely impractical. Solana's architecture addresses these limitations through its Sealevel parallel runtime, which allows multiple non-overlapping transactions to process simultaneously, enabling agent swarms rather than single bots. With block confirmation times of around 400 milliseconds and transaction costs measured in fractions of a penny, Solana makes continuous 24/7 agent operation economically viable in ways Ethereum cannot match. Developers can build real-time market-making agents, multi-step atomic strategies, and coordinated swarms that treat the blockchain as an execution layer rather than a slow settlement layer. Solana does present its own challenges, including fragmented RPC infrastructure, a steeper Rust-based development curve, and careful state management requirements across the account model.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Uses Claude Code to Fix and Upgrade macOS Real-Time Meeting Translation App

A developer building a macOS app that translates Zoom and Google Meet audio in real time discovered the translation stopped after roughly ten minutes due to a 10-minute WebSocket session limit in the Gemini Live API. Anthropic's Claude Code diagnosed the root cause and refactored the codebase to add proactive GoAway signal detection, exponential backoff auto-reconnection, and a 30-second ping keep-alive mechanism. These fixes allowed the app to silently reconnect without users noticing any interruption during long meetings. With stability addressed, the developer then asked Claude Code to suggest new feature directions, spanning UX improvements, translation quality enhancements, and other functional upgrades. The project represents an iterative, AI-assisted development workflow where conversational AI tools handle both debugging and feature planning.

0 comments Read more at DEV Community