Organizational memory that runs on your own hardware
Captures decisions, answers, and institutional knowledge from email, chat, and documents. The knowledge your people already write down, woven into a searchable memory.
You already pay for this. You pay in wasted time, repeated work, and the knowledge that walks out the door.
"What did Lars say about this client before he left?"
Lars wrote it down. In an email. Years ago. It is still in the sent folder, and nobody can find it.
"Has anyone dealt with this regulator before?"
Three people have. Their analysis is sitting in their inboxes and chat threads. You just could not search across them.
"I know we decided this. Where is the rationale?"
The rationale is in a Teams thread. The next person gets the decision, but never the why.
The answer to almost every question is already written down. Finding it is the expensive part.
What it is
No new tools. No extra work. No workflow change.
📥
Passive capture
ThreadWeave reads what people already write, in the tools they already use. Email, chat, documents, meetings. Nothing to fill in, nothing to remember.
🎯
Signal over noise
A two-tier detection engine separates decisions and answers from "ok thanks" and meeting reminders. It knows the difference between a decision and small talk.
🧭
Answers with context
Every captured item keeps its context: who wrote it, when, and for which topic. Search results say why something matters, not just that it matches.
It captures what you would pay to never lose.
The mental model
The palace is the data model
ThreadWeave organizes your memory as a palace: your whole organization, mapped into a structure people already understand.
Palace
The whole organization's memory. Everything captured lives here.
Wing
A team or department. Your structure, mirrored.
Room
A topic inside a wing. One room per subject, no folder sprawl.
Drawer
One verbatim captured answer or decision. The raw source, preserved.
Closet
An LLM-compressed summary of a drawer. The quick version.
Hallway · Tunnel
Connections within a wing, and across wings. Knowledge finds its way between teams.
Wings, rooms, and drawers are real concepts in the product: --wing engineering --room database
How it works
One pipeline, four gates
Every piece of content passes every gate before anything is stored.
Gate 1
Dedup
SHA-256 content hash. Nothing captured twice, even across connectors.
→
Gate 2
Detect
Classifies answer, decision, question, chat, reference. LLM-first, regex fallback.
→
Gate 3
PII gate
Emails, phone numbers, personal IDs detected and stripped before storage.
→
Gate 4
Store
Durable SQLite or PostgreSQL, with MemPalace hybrid search on top.
# Classify a sentence before it is saved
$ threadweave detect "We decided to use PostgreSQL for the auth service"
→ should_save: true, content_type: decision, has_pii: false
Connectors
It plugs in where your team already writes
🏢
Microsoft 365
Email watcher: continuous one-way polling of Exchange Online
SharePoint and OneNote capture, including edits
Teams bot: capture, search, and privacy commands from chat
Graph connector: syncs entries into Copilot and Microsoft Search
🌐
Google Workspace
Gmail: inbox and sent, thread-aware capture
Google Chat: decisions from team spaces
Drive: Docs, Sheets, Slides, PDFs, Markdown
Offboarding harvester for departing employees
🔌
Anything else
IMAP: any provider, no Azure, no COM
Classic Outlook via COM, or exported .eml files
REST API for anything custom
Media: video and audio transcribed on-prem
One-way pull polling. No inbound ports, no webhooks, no tunnels for content.
Retrieval
Search that knows your organization
🔍
Hybrid retrieval
Semantic similarity and keyword search combined, powered by MemPalace. You get the right drawer even when you don't know the exact words.
📊
Four-factor ranking
Semantic match, org proximity, freshness, and author authority. The answer from the person closest to the topic ranks above the lucky guess.
🗓️
Temporal org model
The org chart is time-aware. Query who was on which team in Q2 2023, and the memory follows the people, not the current chart.
🛡️
Confidentiality at read time
Every search is filtered by the seven sensitivity levels before results leave the server. You only see what you are allowed to see.
Sensitivity is detected when content arrives, not assigned by hand. Client-confidential material is only visible to people working on that matter.
🧱
Walled off by default
HR and salary data is locked away automatically. A decision tagged legal-privileged never surfaces outside the people with a right to it.
Audit trail on every read and every delete.
Privacy
The camera sign, built in
Capturing people's work requires their trust. ThreadWeave treats transparency as the selling point, not the compliance cost.
🏠
Runs in your building, not ours
On-prem from day one. No cloud dependency, no data leaving your control. Air-gapped deployment works.
➡️
One-way flow
Email and content move one way, from M365 into your network. Pull polling only. No webhooks, no tunnel relays for content.
📷
The camera sign that talks
An opt-out registry is enforced at ingest by every connector. When knowledge is captured from your content, the Teams bot tells you, personally.
🗑️
Right to delete
Any person can request deletion of captured content about them. Every deletion is audited. Opt-out is respected everywhere, including in the daemons.
Privacy is a feature. People opt in by not opting out, and they always know what is captured.
Operations
Built for people who run production systems
💾
Durable store
SQLite for zero-config on-prem, PostgreSQL for corporate. The palace survives restarts on either backend.
🔄
Daemons that survive restarts
Email and SharePoint watchers run as systemd units or Windows Startup launchers. Crash, restart, resume where it left off.
🧬
Entry versioning
Re-captured documents chain to the original with a version pointer. You see the evolution of a decision, not duplicates.
📈
Metrics and health
Prometheus endpoints, latency percentiles, throughput, dedup hit rate. OpenAPI docs on every endpoint.
🧪
Tested
315+ tests, green. The connector suite is mocked, so the suite runs anywhere, offline.
🚀
Zero-touch releases
CI bumps the version, tags, and publishes on every push. The project ships continuously, in small verified increments.
Status
Live today, shipping continuously
✅
In production use
Email capture, SharePoint and OneNote watchers, Teams bot, Copilot sync, Google Workspace connectors, offboarding harvest, privacy layer, durable store. All live-verified, version 0.4.0.
🔓
Open source, MIT
The whole thing: connectors, detection engine, confidentiality layers. Read it, audit it, extend it. No vendor can take it proprietary.
🏷️
Zero-touch releases
Every code push is auto-bumped, tagged, and published. Small verified increments, live-proven before the next one starts.
🗺️
Next
Whatever the field reports. The backlog is driven by real deployments, not slides. Connector coverage keeps widening.