Agents That Remember — Persistent Memory for Your AI Clients, Powered by Geneziz
Agents That Remember — Persistent Memory for Your AI Clients, Powered by Geneziz
You finish a great session with your AI assistant. It read your bookmarks, answered with citations, summarized three tools. Then you close the window — and it's all gone. Next session, the same assistant greets you like a stranger.
Every "second brain" product on the market solves storage. Almost none solve memory — the part where using it makes the next session better.
Geneziz v1.12.0 closes that loop. Your AI clients don't just read your knowledge base anymore — they leave artifacts in it, and those artifacts compound.
The memory loop, in one paragraph
Geneziz already had the capture side nailed: a browser extension and source connectors (X.com bookmarks, GitHub stars, YouTube, Bluesky, Reddit, Hacker News, plain web saves) that continuously file everything into a local Markdown knowledge base you own. What v1.12.0 adds is the feedback side: a new set of MCP tools that let any AI client ask your knowledge base questions, persist the exchange, and summarize items so the answer is cheaper the second time. Capture feeds memory; memory feeds agents; agents enrich the base. It runs by itself.
The new tools (15 → 33)
The Geneziz MCP server now exposes 33 tools. The four that carry the memory loop:
ask_kb — Ask a question over your entire knowledge base and get an
answer with cited sources (title, type, filename — real provenance, not
vibes). Pass a conversation_id and prior turns are threaded in, so
follow-ups work. Set save: true and the exchange is persisted:
Q: "What did I save about local-first note apps?"
A: [answer] — sources: 3 files cited
→ conversation conv-20260902-195643 saved
Q: "And which of those three works offline?"
(same conversation — the assistant already knows which three)
list_conversations / get_conversation — Your Q&A history is a
first-class citizen. Browse past sessions, newest first; open one and every
turn is there with its sources. Any client, any machine — the memory lives
in your data directory, not in someone's cloud.
append_conversation_turn — Compose or extend conversations manually.
Perfect for logging a decision your team made in a chat window, so the next
agent session starts where the humans left off.
summarize — The cheapest AI write in the product. It generates a 2–3
sentence TL;DR and persists it into the file's tldr: frontmatter.
Re-reading is free — the viewer and every future agent read the stored
summary instead of burning tokens again. Re-running simply replaces it.
And the server grew into a full curation surface: beyond the memory loop,
the 33-tool set covers safe editing at scale — organize_pending
(AI-assisted bulk organization with dry-run previews and an undo journal),
bulk_tag / update_item / refile_item, archive_item / restore_item
with undo_last, add_note, build_wiki (auto cross-links between related
items), detect_gaps / fill_gaps (metadata health), check_url /
enrich_link (link health and rich previews), and find_duplicates /
find_related (embedding-based).
Why "with cited sources" is the whole point
Generic agent-memory plugins give you a blob of remembered context and you
just have to trust it. ask_kb retrieves from your files, shows you
which files, and the answer is only as good as what you actually saved.
That's not a chatbot with a diary — that's retrieval-augmented answers over
a knowledge base you control, with provenance on every claim. When the
answer matters, you check the source. You can always check the source.
Local-first changes the economics
The knowledge base lives on your machine: plain Markdown files, local full-text (FTS5) and semantic (vector) search. The MCP server ships inside the Geneziz app and registers itself into your clients with one command:
geneziz mcp register
Claude Desktop, Claude Code, Cursor, Windsurf, Codex, VS Code, ZCode, and OpenCode are all detected and configured automatically.
Costs are exactly what you'd want: reading the KB (search, listing,
conversations) is free. ask_kb costs 1 credit per turn on Geneziz Cloud —
or nothing at all if you bring your own API key. summarize is the
cheapest write in the system. Your data never leaves your machine unless
you make an AI call.
The habit that compounds
Here's the workflow that changes everything, and it takes zero discipline:
- Keep saving things (you already do — the extension does it for you).
- Ask your assistant questions through
ask_kbwithsave: true. - Occasionally
summarizethe items you touch most.
That's it. Every question leaves a permanent, source-cited answer in your history. Every summary makes a file permanently cheaper to use. A month in, your agents answer questions about your corpus, citing your files, with your accumulated context — something no cloud memory service can offer, because none of them own your capture pipeline.
Your memory feeds itself. Your agents remember.
Geneziz is a local-first AI knowledge engine for macOS and Windows. The MCP server
with all 33 tools plus 4 knowledge:// resources ships in the current release — install, run
geneziz mcp register, and your assistants inherit your memory.
Related Posts
The Geneziz AI Page Grows Up — Streaming Answers, Vision, and Tasks That Run While You Sleep
Geneziz v1.13.0 turns the knowledge base into a living assistant: answers stream in as they're generated, you can attach images to questions, and scheduled tasks interrogate your own corpus on autopilot — then notify you. All local-first, with your choice of model.
Building a Local-First Knowledge Base with SQLite FTS5 and Markdown
A deep dive into how Geneziz stores your knowledge as plain Markdown files with YAML frontmatter, indexes them with SQLite FTS5 for instant full-text search, and why this architecture beats cloud-based alternatives for developers who value data ownership and longevity.