1.26234 — 22 August 2026

Three days after 1.26231, and this one is about knowing things.

If you read only one thing: your gezels can now look things up — in your project, in your linked projects, and in reference libraries you install — and tell you where each answer came from.

Reference libraries your crew can search and cite

A knowledge catalog is a body of reference material packaged as a single file with a .gezk extension: an encyclopedia, a set of manuals, a standards library, your own notes. Install one and it becomes something your gezels — the named AI craftspeople you work with — can search and quote, with the source named every time.

Install them from Settings → Knowledge, by file path, by https:// URL, or with a file picker. Once at least one catalog is registered, a Knowledge area appears in the sidebar: a catalog and topic list on the left, the documents in the middle, the article on the right, with its license, its origin, and a Copy citation button. There is also an Ask a gezel button on any article, so you can go from reading to asking without retyping what you were reading.

Each project decides how much of this it sees. Project settings carry a knowledge row with three choices: all your catalogs, only the ones you pick, or none for this project.

You can build your own. gezel knowledge init and gezel knowledge build turn a folder of Markdown into a catalog, entirely offline — no daemon, no account, nothing uploaded. The whole thing stays on your device, including the search.

Your gezels can look things up now

Until this release, a gezel knew what was in the conversation and what you had told it. If the answer was in a file three folders away, it had to guess that the file existed and go read it.

Now there is a general search tool covering the active project's workspace and artifacts, the project's memory, the gezel's own memory, folder and architecture summaries, and your shared documents library — ranked together, with a citation on every result.

Alongside it, gezel can put a small block of relevant material onto a turn before the model answers, so it starts with the right context instead of hunting for it. Settings → Project knowledge & memory controls how much:

Off — nothing is added. The gezel can still call search itself.
Lean — paths and short hints, about 320 tokens.
Balanced — a diversified set of excerpts, about 1,000 tokens. This is the default.
Deep — broader excerpts including project and area summaries, about 2,800 tokens.

There is an optional token cap if you want a specific number, and gezel applies a second ceiling based on your model's context window, so a small local model is never crowded out of its own answer. That is the reason the modes are worth knowing about: on a memory-constrained model, Lean genuinely answers better than Deep.

Two habits of the feature are worth stating plainly. Retrieved text is labeled untrusted evidence in the prompt — it is material to consider, and it cannot instruct the gezel or override you. And model-facing search is bound to the project you are in: a gezel cannot reach into your other projects, other gezels' memories, or your conversation history to answer a question.

Linking one project to another

Sometimes a project genuinely needs a neighbour's material. Project settings now have a Linked projects list — pick up to 32.

A link is one-way and it does not chain. Linking A to B lets work in A search and read B; it does not give B anything, and it does not reach whatever B is linked to. Linked files show up under a ../project-id/... path so it is always obvious which side a result came from, and the whole arrangement is visible on the one screen where you granted it.

People in your photos

Turn on Recognize people in photos in Settings and gezel looks for faces in the photos it has indexed and groups the same person across them. Name someone once and you can find them anywhere. A People panel appears in each project's overview.

This runs entirely on your machine — the photos and the face data never leave it — and turning it on downloads the face models once, about 260 MB. It is off until you ask, and there is a Turn off and erase face data button that deletes every stored face signature, person, and name across every project. Your photos are not touched by that; only what gezel worked out about them.

Separately, the file browser can now show visually similar images in a project from any image you are previewing.

Every task gets its own folder

Small change, and it removes a whole class of quiet damage.

Work a gezel does for you lands in the project's artifacts drawer. Until now, two runs of the same procedure — a craftbook, gezel's word for a repeatable recipe — wrote to the same filenames and silently overwrote each other's working notes. Two tasks running at once could do it to each other.

Every task now gets artifacts/tasks/<number>/ as its own working folder, created when the task is created, and shared by the shards of a task that fans out into parallel work. Nothing to configure; a task's material simply stays a task's material.

A better first five minutes

Opening gezel for the first time used to mean an empty composer while a multi-gigabyte model downloaded in the background.

The meester — the guildmaster gezel who is your front door — now greets you with a short introduction, an explanation of their role, and a few openers you can press instead of thinking of an opening line. It is a written introduction, clearly presented as one, not words put in the meester's mouth.

And if you send a message while the model is still arriving, gezel now says "43% downloaded — hang tight" instead of telling you to download a model from a list that is not on screen.

Smaller things

Chat reads like people talking. Hand-offs between gezels now render as a short line — who passed what to whom, and about which task — instead of the paragraph of tool-calling instructions the model was actually sent.
Thread previews are readable. The one-line summary under a thread no longer shows raw tool-call markup or hidden reasoning; it says what happened in words.
Tool activity says what it touched, and file paths in it are clickable straight into References.
Word and LibreOffice lock files are hidden. The stray ~$report.docx files that appear while a document is open no longer clutter the file browser.
File references resolve properly. A reference to a specific path no longer falls back to matching some other file with the same name at the root, and old mis-resolved links in existing threads are repaired.
/connect and /disconnect in the CLI hook VS Code, OpenCode, pi, and Codex up to your local models — the same thing Settings → Connected Apps does, from the terminal.
Dates and times are consistent across cards, threads, and panels, and the local-engine setting says "This Mac" or "This PC" rather than a generic device label.
Tasks can be updated in bulk — select several and set their status at once.
Local models re-answer faster. The llama.cpp cache-reuse restriction that cost multi-slot setups their prompt cache is gone, and an MLX bug that made reasoning stay on for turns that had asked for it to be off — inflating both time and tokens — is fixed.
Windows app updates are steadier. Staged runtime cleanup retries rather than failing when a virus scanner or indexer holds a file open, and installer-staged service trees are now verified and reused on Windows the way they already were elsewhere.
The Handboek uses the warm light reading surface when your theme is light, in both document and video modes.
Refreshed catalog content package (gilde 0.1.39).

For developers

.gezk, the knowledge catalog format, is frozen as v1 and implemented in a new @bendyline/gezel-knowledge package: a deterministic compiler, verified archive reading, Ed25519 manifest signing, a Markdown adapter, and a read-only two-stage bit384+int8 retrieval handle. It depends only on core.
New endpoints: /api/knowledge/* for install, enable, quarantine, browse and search; /api/projects/:id/tools/find-similar-images; a people surface for face grouping; and machine-broker knowledge asset endpoints under /v1/remote/manage/knowledge/*. All with client methods.
AI Apps — a project type, crew, craftbooks, scripts, data, and an optional dashboard composed into one reusable experience — have a written guide: Building AI Apps inside Gezel. They package as .gezapp files, and a gezel can export or import one for you.
Retrieval is auditable. Every injected context block records the query hash, policy, estimated tokens, result scores, provenance, and citations as a history event. Raw queries and retrieved text are deliberately not copied into telemetry.
search_code and search_documents are now compatibility aliases for the unified search tool. grep_files remains the right tool for exact strings and regular expressions.
Craftbook steps can scope retrieval, including to knowledge alone, and address their task folder through the reserved {{task.dir}} token.
Embedding instructions are model-aware. The e5 family gets its required query: and passage: prefixes; other families are unchanged. There is a new cross-lingual calibration benchmark with a first Dutch/English baseline.
New architecture notes: per-task artifact folders, the .gezk format, knowledge catalogs, project linking, and project retrieval.

Anything to do?

No. Update when you are offered it.

Two settings are worth a visit. Settings → Project knowledge & memory is where you decide how much your gezels look up on their own; Balanced is a good place to start, and Lean is the better answer on a small local model. Settings → Knowledge is empty until you install a catalog or build one, and it is the interesting half of this release.