Knowledge

The Knowledge pack is for work that has to show where its claims came from. Wordcell keeps notes beside code where agents can find them, Oh stores memory as typed records with sources, and Sponge runs research on your own machine. Hand research and recall to an agent, and you can still check where each answer came from.

Why typed memory

Oh describes memory built on a vector store as one where facts go in, answers come out, and nobody can say why. Each Oh record instead declares a kind, a stable key, the keys it depends on, and canonical JSON content, so equal content always has the same SHA-256, and every fact keeps its sources and a history that replays. A versioned ontology separates a claim from someone’s stance toward it, and evidence from the assertion it bears on.

Rules can derive new facts with proofs, and a derived fact is never written back on its own; saving one takes a reviewed write. Agents get a working store the host can purge, kept apart from reviewed memory, and every answer says which store each premise came from. Oh checks shape and references. It cannot tell whether a statement is true, and says so.

Products

Wordcell

Markdown notes kept beside your code, such as decisions, plans, and sources, that coding agents can find. Exact search needs no account or model.

Oh

Open-source agent memory that stores each fact with its sources and a replayable history, in one SQLite file at .oh/oh.sqlite. oh verify replays the operation log, and a Rust (WebAssembly) encoder and a TypeScript encoder are tested to agree byte for byte. It is MIT licensed.

Sponge

A local research harness: a background service with a terminal UI and a macOS menu-bar app, offering brief research and a deeper mode. It uses ALGAL to evolve its brief-research strategy, xcb for all model access, Wordcell for documents and publishing, and GhostGet for signed-in capture. Its command is spongev2. It is in development and not yet released; it runs from a private checkout. The hosted private library at sponge.computer is the earlier Sponge product.

The literature project in Bio runs on the harness. A daily GitHub Actions job watches for new papers, writes a digest and publishes it, and the output feeds the papers page. Its measured gains came from engineering changes checked by backtest, not from ALGAL evolution: a scorer rewrite raised recall at 25 papers a week from 28% to 47% on 211 held-out papers, and a cheap screening model raised recall at 50 a week from 55.9% to 62.6%.

What Sponge relies on

ALGAL runs agent programs that wait for approval, resume after a crash, and replay from a record of each step. GhostGet gives agents reviewed web actions and holds no password. Excalibur (xcb) routes model work across subscriptions you already pay for; it and ALGAL are covered in Tools for Token Maximalists.

How they connect

Wordcell → Oh: proves

Wordcell loads notes into a disposable Oh graph and answers six named queries: backlinks, reachability, relation closure, scope route, shared tags, and shared concepts. Each answer carries a proof: the source note, a fingerprint of that note’s text, and the fingerprint of the Oh record built from it. When Wordcell says a retry plan depends on a parser rule, the line that makes the link can be read. Edit the note, even its spacing, and the fingerprint changes, so old proofs go stale instead of quietly standing. Wordcell uses Oh only for these provable graph queries. Its search and agent memory do not use Oh, and deleting the graph file loses nothing. Graph queries accept up to 4,000 notes, 100,000 facts, and 64 MiB of text, and a truncated result exits with code 4.

Sponge → Wordcell: publishes

The harness extracts PDF text with Wordcell’s pipeline, renders a static Wordcell edition with publish, and puts selected documents on a wordcell.io site with hosted publish.

Sponge → ALGAL: evolves

Brief research follows a strategy: a small ALGAL program with four settings (guidance text up to 1 KB, a source limit, an excerpt size, and whether to look for counterevidence) and no agent calls of its own. evolve propose asks the model for one candidate. evolve evaluate runs the current strategy and the candidate on frozen cases, two visible during the proposal and two held out, and the scorer requires exact finding sentences, exact cited source sets, and the stated uncertainty. A candidate passes only if it improves with no regression and passes every holdout, and the host signs the report.

On projects created with evolution turned on, evolve adopt replays the whole evaluation inside ALGAL before committing a revision that changes only the strategy, and evolve restore rolls back through the same gate. Nothing is adopted automatically, deep research keeps a fixed policy, and any other commit between evaluation and adoption makes the recommendation stale.

Sponge → GhostGet: captures

All signed-in capture goes through GhostGet, paced per login. The harness stores only a login name, never cookies or credentials.

Sponge → Excalibur (xcb): routes

Every model call goes through xcb, run from a hash-pinned executable. Provider sign-ins stay in xcb, and research draws on coding subscriptions the owner already pays for.

Excalibur (xcb) → Wordcell: searches

xcb can bind one Wordcell vault so its workers search it read-only and cite what they find.