Every coding agent starts each session knowing nothing, and most of the friction of working with one is re-explaining what you already figured out. A batch of recent Show HN launches attacks that from several angles: search over your own past sessions, a shared memory that spans harnesses, conflict-free sync between machines, and the administrative tasks nobody wants to do by hand.
Search the sessions you already have
Skillsync takes the position that agent memory should just be search. It is a macOS app that checks every prompt against every past session your agents have already written to disk, and returns a short recollection with sources in under a second. There is nothing to index and nothing to configure: it installs plugins for Claude Code, Codex, Cursor, OpenCode, Amp, pi, Grok, Antigravity and Hermes, reads the sessions they keep, and gets out of the way when the past is not relevant. A shortcut opens search over any app, so you can find and reopen a session without leaving what you were doing.
One memory for every harness
SuperLocalMemory is a local-first memory control plane that aims to serve every agent you use at once, from Claude Code and Cursor to web bots that speak MCP, with decisions, documents, screenshots and PDFs recalled the way you ask for them. It is governed like company data rather than like a scratchpad, local agents and web bots talk to each other through a mesh, and it has a stated rule about honesty: when the answer is not in memory, the agent hears that it is not there rather than a confident guess. That refusal is the feature most memory layers skip.
Memory that converges across machines
Sharing one memory between several computers usually means running a server. Server Noonien avoids that. It is a drop-in replacement for the official MCP memory knowledge graph, keeping the same entities, observations and relations, but it converges across machines through peer-to-peer sync with no central server, no database and no merge conflicts. A `nooniend` daemon replicates each node's shard straight to the others, so several machines share one memory over a VPN or LAN with no shared folder; where peers cannot reach each other, file and S3 backends cover the same ground.
Ranking that stays on your machine
The same local-first instinct shows up outside agents. Okaeri is a localhost homepage that gathers papers, blogs, forums and news in science and tech into tabs you set up yourself. A small embedding model runs on your own machine to rank stories, your clicks and hides feed the ranking, and your reading history never leaves the device. The ranking algorithm is yours to tune, from the UI or in the code. The trade-off is stated plainly in the README: about 500 MB of disk, a model download on first run, and a background process that briefly uses a couple of CPU cores and some battery.
The library you actually keep
Linkit is a desktop bookmark manager that treats saved links as assets rather than clutter. AI summaries, semantic search and a knowledge graph sit on top of your bookmarks, collections group ideas across categories, and link health checks and duplicate detection keep the library clean. One-click import covers Chrome, Firefox, Raindrop and more, storage is local-first with optional cloud sync, and the AI features can run against any OpenAI-compatible API or a local model such as Ollama. It is free, on macOS, Windows and Linux.
The paperwork nobody wants
Two launches handle the administrative side. Costura is open-source, self-hosted AI spend attribution. You connect read-only billing keys for Anthropic, Claude.ai, OpenAI, Codex, OpenRouter, Cursor and GitHub Copilot, and it folds them into one daily ledger rolled up to the person who spent it. Anomalies come back as a sentence naming the key and model that caused them, keys with no owner or no limit are flagged, and the ledger is tied out against each provider's own report. Provider keys are encrypted on your infrastructure and never leave it, and the tool never creates, edits or deletes anything.
Claude Migrate solves a smaller and more annoying problem. Switching Claude accounts on a Mac leaves the Desktop Code tab empty, because Claude Desktop keeps one sidebar store per account and organisation and only reads the one you are signed in to. This unofficial CLI copies the session records from the old store to the new one, keeps their IDs so pins survive, clears the links that belong to the old account, and never touches transcripts. It dry-runs first and is undoable. It is macOS only, and it depends on undocumented storage that can change.
Keeping up without a feed
Standard Topic rounds the batch out on the reading side: a news platform that ranks the latest reporting on every topic by relevance and adds AI summaries that show their sources, covering world, politics, business, technology and AI, plus a twice-daily podcast. The useful property is that the summaries point back to the reporting rather than standing in for it.
Why this batch matters
Agents do not need to remember everything. They need to be able to find the one decision from three weeks ago, in the harness you happen to be using today, without you retyping it.
