Skip to content

InsightMesh Core

A cognitive knowledge engine that compounds understanding over time through multi-agent investigative inquiry, persisted as an evolving Obsidian wiki.

InsightMesh turns your AI chat history into a growing wiki you actually own. Local-first, cross-linked, transparent about what it knows.

Latest: Spec 005 — per-page provenance

Every wiki page now carries a provenance: frontmatter block (latest checkpoint, contributing conversations, action, confidence, total edits, cumulative exchange count) backed by a permanent per-checkpoint JSON record at <vault>/InsightMesh/.history/checkpoints/<conv-id>/cp-<NNN>.json plus a shadow git repository for git log -p style diff history of every page's evolution across edits. All three transcript shapes (Claude.ai, ChatGPT, Spec 001 flat-array) validated on real data. See the CLI reference for the on-disk layout and specs/005-page-provenance/quickstart.md in the repo for jq + git log recipes. A dedicated in-Obsidian viewer plugin (insightmesh-obsidian) is now available in the Obsidian community plugin browser: in Obsidian, open Settings → Community plugins → Browse, search for "InsightMesh Viewer", and Install. Renders the active page's provenance in a side pane with click-through to its source conversation, checkpoint history, and snapshot-to-snapshot diffs. Compatibility: viewer 0.x.x supports core schema_version=1. Pre-release builds available via BRAT (Add Beta Plugin → aucontraire/insightmesh-obsidian). File issues at aucontraire/insightmesh-obsidian/issues. Prior milestones still apply: Spec 004 (long-chat checkpointing with --resume / --max-exchanges), Spec 003 (attachment + pasted-text synthesis), Spec 002 (insightmesh list + multi-conversation export selection via echomine).

What it does

You spend hours having intellectually rich conversations with Claude or ChatGPT — and then lose all of that context the moment the session ends. InsightMesh fixes that by reading your transcripts and synthesizing them into organized, cross-linked Obsidian wiki pages.

  • Sub-agent pipeline: Synthesis → Historian → Editor, each with a single responsibility, all coordinated via the Claude Agent SDK.
  • Cross-linking that compounds: the Historian searches your vault (via MCPVault) and weaves [[wiki links]] between new and existing pages.
  • Honest reasoning trail: every page write decision is logged with full rationale — what signals matched, what got skipped, why.
  • Local-first: data stays on your machine in your Obsidian vault. No cloud, no accounts.

How it differs from NotebookLM, Perplexity, etc.

  • Knowledge compounds: not a one-shot research tool — inquiry #50 is richer than inquiry #1 because prior pages get pulled into the synthesis.
  • You own the data: markdown files in your Obsidian vault, version-controlled, portable.
  • Intellectual transparency: the multi-agent process is visible in the output. Every decision has a rationale.

Where to go next

  • Getting Started

    Full beginner walkthrough: prerequisites, install, Obsidian vault setup, Claude Code plugin install, smoke test, your first real chat.

  • Known Limitations

    Honest list of what doesn't work yet, what's slow, and what's planned next.

Architecture (Phase A)

flowchart TD
    A[chat transcript JSON] --> B[Synthesis Agent]
    B -->|drafts| C[Historian Agent]
    C -->|augmented drafts| D[Editor Agent]
    D -->|EditorOutput| E[Obsidian vault + session log]
    F[(MCPVault)] -.->|BM25 search| C
    F -.->|write or patch| D
    G[(obsidian-skills plugin)] -.->|preloaded| B
    G -.->|preloaded| C
    G -.->|preloaded| D

Three sub-agents defined as markdown files in .claude/agents/, orchestrated through claude-agent-sdk. The Editor uses the kepano/obsidian-skills obsidian-markdown skill for proper wikilink and frontmatter syntax.

Phase B will eventually migrate orchestration to LangGraph for deterministic execution — but that is triggered by orchestration needs, not the next step. The trigger is the first feature that needs reliable parallel fan-out/fan-in (the multi-perspective Critic + Researcher agents), or the point at which the SC-001 timing limitation and orchestrator overhead become a real constraint. Until then the roadmap stays on the sub-agent pipeline: adding agents and richer content handling (attachments, images) as sub-agents while the prompts and JSON contracts stabilize.

Status

Feature Status
Chat-to-wiki batch synthesis Spec 001 — working
Multi-page cross-linking
Session logging + decision rationale
Same-topic update detection
Multi-conversation export selection (pick a chat from a Claude.ai/ChatGPT export) Spec 002 — working
Pre-flight validation (vault + agent presence checks) Spec 002 — working
Attachment and pasted-text synthesis (Claude exports) Spec 003 — working
Long-chat checkpointing + auto-resume + per-invocation cap Spec 004 — working
Per-page provenance (checkpoint JSON + frontmatter + shadow-git diff history) Spec 005 — working
Live inquiry (ask questions, refine, synthesize) planned
Bias/assumption checking (Critic agent) planned
Web research (Researcher agent) planned