Crystals: A Push-Based Agent Memory System That Delivers Notes Before Actions Run

·3 views

Developers at DEV Community have introduced a concept called 'crystals' — a push-based agent memory mechanism that injects relevant notes into an AI agent's context immediately before a specific action executes, rather than waiting for the agent to query for help. Unlike traditional retrieval-augmented systems where memory is only fetched when the agent suspects it needs guidance, crystals are bound to particular actions and fire automatically based on pattern matching. Each crystal is a markdown file containing a concise, single-piece instruction delivered between defined markers, keeping the payload small and precise to avoid context-window budget issues. The approach addresses a documented gap in agent memory research: a failure mode where a correctly matched memory note is silently dropped by a token-budget packer before ever reaching the model. The authors frame this as a practical field report responding to a 2026 survey that identified bounded-memory management as an open problem in agent memory systems.

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)

How Digital Footprint Tools Read Your Data and What Stays on Your Device

Digital footprint check tools analyze exported social media archives to generate privacy risk scores, drawing primarily on post history, direct messages, IP records, and linked contact details. The core scoring relies on pattern matching within tweet history to flag phone numbers, emails, addresses, and sensitive topics, while other files serve as cross-checks to strengthen or contextualize findings. Certain data categories — such as ad engagement records, follower lists, and device tokens — are deliberately excluded to avoid inflating results or involving third-party data. All parsing, file reading, and scoring processes run locally on the user's device, with post content never transmitted over the network; only post deletion requires an authorized API connection. Discrepancies between expected and reported item counts are typically explained by the archive's snapshot date or the inclusion of previously deleted content.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds AI-powered gossip tabloid for RimWorld colony that must cite its sources

A developer used Claude Code to create 'The Estian Tattler,' a fictional tabloid that generates gossip stories drawn entirely from a RimWorld game save file spanning roughly 300 in-game days. The project ingested 631 records — including marriages, conversations, letters, and game messages — from a modded colony called the Tribe of Estian. Every sentence the AI reporter writes must cite specific records as evidence, and a plain TypeScript fact-checker automatically rejects any claim that names a colonist or includes a number not supported by the cited sources. Failed drafts are returned to the reporter, with stories spiked after three failed attempts, while passing drafts await human editorial approval before publication. Six editions have been published so far, with printed stories drawing on 136 of the 631 available records, and a dashboard tracks which colonists the paper has yet to cover.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

PPO, GRPO, DAPO, GDPO Compared for Training LLM Agents on Multi-Reward Tasks

Training large language model agents for real-world enterprise tasks requires balancing multiple competing reward signals — such as task success, execution efficiency, and safety constraints — rather than a single scalar score. Simply combining these reward channels into standard algorithms like PPO or vanilla GRPO often causes training instability, where high-variance rewards drown out subtler signals. PPO's actor-critic architecture also imposes significant hardware costs, requiring a second model of equal size to serve as the critic. Newer methods like GRPO, DAPO, and GDPO attempt to address these issues through group-relative advantage estimation and decoupled channel normalization. An empirical study compares seven trainer configurations on Qwen3-14B across these approaches to evaluate their effectiveness on multi-reward agent post-training.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds persistent memory system for Mac AI coding app Deiko

A developer building Deiko, a Mac app that captures screen context and voice input for AI coding agents, grew frustrated that agents had no memory of previous debugging sessions on the same issue. After noticing repeated explanations of identical bugs across days, he spent a week engineering a memory layer to let Deiko recognize when new briefs relate to past tasks. He evaluated Jev, a probability-focused model from TypeSafe AI priced at $0.042 per million input tokens, which returns a confidence score on whether two items are related rather than a verbose response. When TypeSafe paused signups, he rerouted through gateways including Vercel AI Gateway, OpenRouter, and Cloudflare to access the same model. Drawing on how tools like Claude Code, Mem0, and Sentry handle memory and issue grouping, he built a small filing algorithm that combines exact contextual clues with model-based classification to cluster related briefs automatically.

0 comments Read more at DEV Community