What AI Agent 'Memory' Really Is: A Technical Breakdown

·2 views

A DEV Community analysis argues that AI agents have no true memory — only a context window that is reconstructed from scratch before every API call. The article identifies three functional memory patterns: a live transcript, a fact store (essentially a database), and external state such as files or task lists. The author contends that borrowing terms like 'short-term' and 'long-term' memory from cognitive psychology is misleading, since no automatic consolidation between stores exists. For long-running agents, the piece recommends deliberate compaction policies — such as eliding old tool outputs and summarising middle reasoning — rather than simple truncation. It also highlights plain external files like a plan.md as an underrated, inspectable alternative to vector databases for maintaining durable agent state.

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)

Interpolation, GANs or Diffusion: Choosing the Right Image Upscaling Method

Every image upscaling method works by generating new pixel data that was never in the original image, since resolution limits mean lost detail is permanently unrecoverable. Interpolation methods like bicubic and Lanczos are fast and deterministic, inventing only smooth transitions but producing soft or ringed results. Learned feed-forward models such as ESRGAN invent textures based on their training data, but fail when the input degradation does not match what they were trained to reverse. Diffusion-based refiners go furthest, generating entirely new content guided by a prompt, offering the most visually convincing output at the cost of faithfulness to the source. The right choice depends on whether the output must serve as evidence of the original input or simply needs to look plausible.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Image dimensions, not file size, determine AI vision API token costs

When sending images to AI vision APIs, the number of input tokens charged depends entirely on an image's pixel dimensions, not its file size or compression level. This is because providers decode images to a raw pixel grid before processing, meaning a 40 KB JPEG and a 4 MB PNG of identical dimensions cost exactly the same. Each major provider uses its own formula: OpenAI tiles images into 512×512 patches after resizing, Anthropic applies a straightforward area-divided-by-750 calculation, and Google uses a flat 258-token tile system with a sharp cost jump above 384 pixels per side. The most effective way to reduce costs is resizing images before upload, since compressing files more aggressively only saves bandwidth, not tokens. These formulas were accurate as of August 2026 based on each provider's own documentation, but should be verified before use as rates and rules can change.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo Developer Builds Vinyl and Hi-Fi Community Platform Using AI Tools

A solo developer has launched El Club del Vinilo, a Spanish-language website combining a vinyl and Hi-Fi magazine, music discovery features, and a user community for sharing collections and setups. The platform includes a section called Café Concierto, which organises hundreds of live concert videos by genre, and a floating music player that lets visitors listen without leaving the page. The developer used AI tools including Claude to assist with coding, debugging, and feature design throughout the build process. While AI accelerated development significantly, the creator notes that core decisions about product direction and user experience still required human judgment. The project is presented as a case study in how much a single developer can now build independently using modern AI-assisted development tools.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Image APIs Block Harmless Requests and How to Diagnose Refusals

AI image generation APIs use multiple independent filtering layers — including prompt classification, input image screening, and output classification — each of which can block a request for different reasons. Latency is a useful free diagnostic, since an instant refusal points to a text classifier while a slow one suggests the image was generated before being flagged. Certain benign content categories, such as medical illustrations, art history references, kitchen tools, and historical imagery, are frequent false-positive triggers because they sit close to restricted content in a classifier's feature space. Developers are advised to log refusal times, error codes, and category distributions to better understand patterns specific to their product. Importantly, if a request genuinely violates a provider's policy, finding alternate phrasing to bypass filters still constitutes a terms-of-service violation.

0 comments Read more at DEV Community