thinkweave

Thinkweave

A self-maintaining knowledge layer for Claude Code — it actively gathers what you read, captures what you do, and curates both into living markdown you own.

Most "agent memory" is a passive store: it remembers what you told it and hands it back. Thinkweave takes the opposite posture. It goes and acquires knowledge on the topics you steer it toward, synthesizes your sessions and your sources into structured pages, maintains its own vocabulary every night, and tracks which of its memories actually got used — all in plain Obsidian-native markdown. SQLite is a throwaway index, rebuildable from the vault at any time.

Markdown is the source of truth · the vault acquires and curates itself · MCP tools + hooks are the interface · nothing is trapped in a store you can't read.

It ships as a Claude Code plugin: weave_* MCP tools, slash-command skills, subagent workers, and four hooks that auto-capture session events.

Why Thinkweave

Retrieval over a pile of notes is table stakes — every memory tool does it, and plenty do it well. Thinkweave's bet is on the three things around retrieval that almost none of them do.

1. It actively acquires knowledge — steered by you, informed by your work

Thinkweave doesn't wait to be told what to remember. The discover → drain spine runs a configurable strategy list that finds gaps and enqueues external sources — papers, repos, articles, news, podcasts, newsletters, YouTube — then fans out subagent workers to write each one up. Two forces aim it:

• Your priorities. Focus areas and intake rails (RSS feeds, mail labels, channel allowlists) live in PRIORITIES.yaml / sources.yaml — data you edit, not code you fork. You decide what flows in, and adding a source type is a registry entry, not a subclass.
• What you're actually doing. The exploratory questions you ask get classified and tallied into probe pressure per concept; the nightly priority worker reads that pressure, checks whether the vault already covers the angle you keep probing, and enqueues fresh acquisition where it doesn't.

The result is a knowledge base that grows toward your work — without you hand-maintaining a reading list.

2. It maintains itself — and hands you a report

Every night the /dream cycle runs ten subagent workers in two phases. It mints concept-hub pages and theme arcs from the day's material, dedups and coarsens the ontology for you (cosine drift-v2 + remembered verdicts), promotes proven proposed-concepts to canonical, reconciles Claude Code's own auto-memory against the vault, and judges yesterday's predictions against what actually happened. You wake up to fresh knowledge digests (vault/digests/) and a maintenance log (vault/.weave/maintenance.jsonl) of exactly what changed — the vocabulary curates itself instead of rotting, and you can audit every move.

3. It learns which memories were worth keeping

Thinkweave snapshots what actually got retrieved: every retrieval MCP call is logged as a context-served event. Decisions are evidence-gated — superseded only lands when git-blame proves the predecessor's lines were really replaced, not merely declared dead — and a decision can carry a predicted_outcome: that gets judged later against reality, with weave rlvr export shipping the substrate for reward modeling. It's memory with a feedback loop, not a write-only log.

…and the foundation under all three

• Internal + external, one ontology you control. Sessions and decisions (what you do) and sources and concepts (what you read) are the same primitives under one shared vocabulary — so a finding from a paper surfaces against an unrelated project when they cite the same concept. Retrieval (FTS5, embedding similarity, typed graph walk, plus budgeted compositions) is the access mechanism here, not the headline.
• Yours, in the open. Human-readable markdown in an Obsidian-native vault — git-friendly, browsable, hand-editable. The SQLite index is throwaway: delete it and weave index --full rebuilds it from the markdown.
• Steerable, not opinionated. Ontology, themes, source types, and priorities are data-shaped extension surfaces (YAML registries, markdown frontmatter) — no plugin classes to subclass, no baked-in schema; the shipped ontology is a minimal seed that grows with your vault.
• Lightweight and local. No always-on server eating RAM. SQLite index, optional embeddings (retrieval degrades gracefully to keyword search without an API key), MCP server launches on demand.

See ARCHITECTURE.md for how the pieces fit, and CLAUDE.md for the in-session agent contract.

How it works