A coding agent is only as good as the three things underneath it: what it remembers, what it can query, and whether anyone is counting what it spends. The newest batch of Show HN projects is unusually concentrated on that layer, and the designs are more conservative than the hype suggests. Where 2025 reached for a vector database and hoped, these tools reach for deterministic state, plain SQL and explicit budgets.
Memory that knows what is current
Retrieval has a known failure mode. Ask for a user's city and similarity search returns both "I live in Delhi" and "I moved to Bangalore", leaving the model to guess which one still holds. A conversation is a timeline, and a bag of embeddings is not a timeline.
TenseMem is built around that single complaint. It is embedded, deterministic state memory: one SQLite file, no server, where the model proposes typed facts and the engine decides what is current. It keeps one current value per slot, reports whether a read is fresh, and refuses to learn from its own output. You install it with pip, and the bundled rule-based extractor means the whole API works without an API key before you wire in OpenAI, Anthropic, Ollama, LiteLLM or Gemini. The tradeoff is deliberate: it replaces fuzzy recall with a smaller set of facts you can actually depend on.
Storage an agent can query directly
The second layer is where the structured and semantic data live. Running Postgres, a vector store and a cache, then writing glue code so the agent can reach all three, is a lot of machinery for a personal tool.
Inillucent collapses that. It is an embedded SQL database in Rust built as an agentic database: a complete SQLite dialect with a Postgres dialect listed as upcoming, vector embeddings for semantic search, keyword search, and a command line, a shell and an MCP server in the box. The MCP server is the point. An agent can query relational and semantic data through one interface without anyone writing an adapter layer. As with any vendor benchmark, read the published speed numbers as the vendor's own measured workloads rather than a result you will reproduce.
Inference across the machines you already own
Not every model fits in one computer's memory, and not everyone wants to rent a cluster.
Lumabri runs a private inference service across your own computers and servers, powered by Colibri and written in pure C. You choose the models, select which computers may help, and approve how their memory is used. The clever detail is that a worker does not need a complete copy of the model: in the resident path it loads the weights for its assigned layers plus whatever shared state the adapter needs, and conversation state stays with the nodes executing those layers. The project is explicit that adding computers adds usable memory and capacity but does not guarantee more tokens per second, which is a refreshingly honest caveat.
Figuring out what fits at all is its own problem, and RigCheck answers it directly. You pick a GPU or a Mac plus your RAM, and each model's checker answers from the requirements and benchmarks published by the model developers, covering language, image and video routes with workflow files and sourced memory limits. It compares specific quantizations rather than giving one verdict, and it says plainly that the numbers are capacity estimates for llama.cpp at a 32K context rather than measured runs.
Instructions and a browser of their own
Two more pieces of substrate showed up this month, both aimed at what the agent is allowed to do.
Skillstore is a searchable skills library for Claude Code and Codex. You connect through MCP to find expert instructions for whatever you are building, publish your own team's methods so agents and teammates follow the same standards, and improve skills with feedback. The public library needs no Skillstore account, so you keep the AI account you already have. Sample skills cover things a reviewer would actually ask for: checking access to each resource rather than just sign-in, validating untrusted input at the boundary, and writing an abuse case and testing it.
Ego Lite is a browser built for agents to run automation in parallel while you keep browsing. Every agent gets its own fully isolated Space, so background work never collides with your tabs, and agents write JavaScript functions against the page instead of getting stuck in command loops. It is MIT licensed and free to download for macOS, with Windows and Linux listed as coming soon.
Two meters on the same problem
The last two tools are both instruments, and they measure different things.
Agent Usage Guard is a Claude Code mod that acts before money is spent. It reads your plan's live usage, counts agents across every session, and holds subagent fan-out, heavy-context churn and retry loops by asking in Claude Code's own dialog. The distinction that matters is timing: a dashboard tells you the window is gone, while a guard asks before it goes.
XTrace is the dashboard. It is an open-source macOS app that reads the Claude Code, Codex and Cursor history already on your Mac and reports agent activity, estimated human time and API-equivalent spend, with Claude and Codex account usage and pace in the sidebar. Scoring skills are customizable, so you can decide what counts as useful work. It is honest that the human-time and spend figures are estimates rather than billing data.
Run them together and you get the thing most teams are missing: a number for what an agent costs you, and a check that stops the number before it happens.
