Changelog

Proactive discovery without silent conclusions

KnowMine can identify preferences, working patterns, and stable facts that may be useful across future AI conversations. Each discovery first enters the Memory Inbox. It does not alter your Soul profile or become an authoritative memory until you approve it.

Every suggestion explains:

• What the AI proposes to remember
• Which conversation evidence supports it
• Where the memory will help after approval

Three clear choices

• Add to my profile: confirm the durable insight
• Edit and add: keep the useful part in your own words
• Decline: mark it inaccurate, context-only, or outdated

Declining a new candidate never deletes an earlier confirmed insight. Revisions retain their relationship to the prior version, reducing one-conversation misinterpretations.

From “saved” to “actually useful”

The home page now shows pending reviews, confirmed traits, and verified successful recall_memory calls. Based on your current state, KnowMine recommends one meaningful next step: review, connect an AI, complete the first recall, or inspect your profile.

Updates that lead to the feature

The in-app banner and MCP notice now share one release catalog. Each update points directly to the relevant experience instead of sending every user to a generic changelog page.

Search Engine Upgrade: pg_trgm + RRF Fusion

Keyword search upgraded from ILIKE pattern matching to pg_trgm trigram matching with native Chinese support and real relevance scoring (0-1).

Results now use Reciprocal Rank Fusion (RRF) to merge vector semantic search and keyword exact search — entries found by both paths rank highest.

What you'll notice: searching for names, product codes, and technical terms is more accurate; Chinese queries no longer miss matching entries.

Knowledge Revision History (Compiled Truth + Timeline)

Every time you update knowledge via update_knowledge, KnowMine automatically saves the previous version.

New MCP tool get_knowledge_history lets you:

• See how a piece of knowledge evolved from first draft to current version
• Trace who made each change (user / AI Agent)
• Current content = Compiled Truth (always latest), history = Timeline (append-only)

Explicit Knowledge Links (Typed Links)

Beyond auto-discovered vector similarity, you can now tag semantic relationships between knowledge entries:

Type	Meaning	Example
based_on	Based on	"New analysis" based on "old research"
refutes	Refutes	"New data" overturns "old conclusion"
extends	Extends	"Follow-up" extends "original report"
evolved_from	Evolved from	"v2 plan" evolved from "v1 plan"
related_to	Related	General association

New MCP tools:

• add_knowledge_link — Create semantic links
• get_knowledge_links — View all explicit links

MCP Tool Summary

Change	Count
Total tools	16 → 19
readonly tier	8 → 10
readwrite tier	15 → 18
MCP Server version	1.5.0 → 1.6.0

All API keys auto-upgrade, no action needed.

Claude Code Plugin

Connect KnowMine to Claude Code in one line:

/plugin marketplace add github:YIING99/knowmine-claude-plugin

What's new:

• Zero-config install: Claude Code prompts for your API Key, stores it in the system keychain, and auto-registers the MCP server
• All 14 MCP tools ready out of the box: add_knowledge, save_memory, search_my_knowledge, recall_memory, get_soul, and more
• Built-in conversation-distill skill: structured session wrap-up — classifies outputs into 6 categories, confirms with you, then writes to KnowMine

conversation-distill standalone is also live on ClawHub:

npx clawhub@latest install conversation-distill

Works with any MCP notes tool, or outputs clean Markdown when no tool is configured.

Get your API Key: knowmine.ai/settings/mcp

MCP Settings Page Update

The Settings → MCP Connections page now features a Claude Code Plugin install guide pinned at the top, with a one-click copy command. The Claude Code CLI section title has been updated to clarify the difference between Plugin install and manual CLI setup.

Make every knowledge save visible inside the AI conversation

KnowMine's core scenario is simple: while chatting in Claude / ChatGPT / Cursor, you use MCP tools to save valuable judgments from the conversation directly into your knowledge base.

But until this release, saving gave you zero immediate growth feedback — you had to open the web dashboard to feel the knowledge base growing. This update closes the last mile of the reward loop.

What you'll see now

After saving a memory, your AI tool will naturally tell you:

Memory saved → folder: Personal Learning Notes
🔗 Auto-linked to 2 related items:
  · Karpathy LLM Wiki Ecosystem Overview
  · KnowMine Reward System: Five Core Decisions
💡 This brings your "Karpathy ecosystem" topic to 5 notes — want me to draft a quick synthesis?

The last line is the key — it's not a stat for you to read, it's a conversation cue for the AI. Claude sees it and naturally takes the bait: "Sure, let me pull them together." It then actually calls get_related_knowledge, stitches those 5 notes together, and produces a real, usable topic synthesis.

Every save becomes a potential trigger for knowledge reuse.

Why no streaks

We deliberately did not implement "3rd today / 12-day streak" style feedback. Reasons:

• Streaks encourage formal low-quality check-ins, but a knowledge base needs high-quality content to compound
• Fixed numerical feedback decays into noise — your brain filters it out within two weeks
• More importantly: the AI repeats streak phrases in follow-up turns, and hearing the same robotic line for 30 days straight is cringe

We took the more restrained but more alive path: topic-cluster awareness + AI action hook. Full design walkthrough in the blog post Why KnowMine's Knowledge-Base Reward System Rejected Streak Counters.

Trigger conditions

• The saved item's tag has ≥ 3 same-tag notes in your knowledge base
• Three tone tiers at 3 / 5 / 10 count, guiding you toward "view relations / draft synthesis / outline an article"
• Never fires on: items without tags, edge-case topics, or the duplicate-reinforcement path — follows the "don't-disturb" principle

MCP tools affected

• save_memory — AI memory capture
• add_knowledge — general knowledge entry

Voice notes, document import, and other write paths will plug into the same reward hook in future releases.

Documents are no longer dead files

Before this release, dropping a PDF into KnowMine gave you an "attachment" — you could search the filename but not the judgments inside. This update connects the document layer and the knowledge layer end-to-end.

What you can do now

1. Upload, auto-parse Supports PDF / Word / Markdown / plain text. After upload, AI will:

• Extract key judgments, conclusions, and action items
• Tag and auto-classify
• Vectorize for semantic search

2. One-click save to knowledge base Every extracted judgment can be individually selected and saved as a knowledge item — no manual copy-paste.

3. Evidence-chain traceback Each knowledge item extracted from a document keeps a reference back to the source paragraph. Open the knowledge detail page and you can jump back to "this came from page X paragraph Y." AI extraction is no longer a black box.

4. Directly readable by MCP tools Once ingested, documents become searchable through search_my_knowledge and recall_memory in your AI conversations — the document layer and the AI-chat knowledge layer are one stream.

Why the evidence chain matters

The biggest risk with LLM document extraction is "you don't know whether it hallucinated." With a reference back to the original paragraph:

• You can verify any AI-extracted judgment in one second
• Six months later, you can still recover the full context
• If the source document is deleted or updated, the knowledge item doesn't become an orphan

Where to find it

• New "Documents" entry in the left navigation
• Document detail page with inline preview + AI analysis results
• "Save as knowledge" button on each extracted insight
• Ingested knowledge items show "from Document X, paragraph Y" on their detail page

Giving the final word on Soul back to you

Soul is the cognitive profile KnowMine builds from your knowledge base. Previously it was entirely AI-written — you could only read it, not change it. If AI got a detail wrong or you wanted to emphasize something, your only option was to wait for the next regeneration.

This release turns Soul into an editable document you and AI collaborate on.

What you can do now

1. Manual editing The Soul detail page now has an Edit button. Inside is a plain text editor — rewrite any paragraph, add details, or delete things that aren't accurate.

2. Every change goes into version history Each save produces a version snapshot. The version list shows:

• Who made the change (you / AI regeneration / accepted AI suggestion)
• What changed
• One-click rollback to any previous version

No more fear of "I broke it and can't get back."

3. AI Suggestion Mode If you don't want to rewrite from scratch but also don't want AI to overwrite everything in one shot — switch to AI Suggestion Mode. AI proposes edits in the margin, you accept or reject each one, code-review style.

Design intent

Soul is deeply personal. It should be:

• Assisted by AI, not defined by AI
• Correctable, with a transparent correction path (version history)
• Clear about what came from you vs. what came from AI — not a blur

After this release, the final word on Soul is yours — AI proposes, you decide.

Where to find it

• "Edit" button in the top-right of the Soul detail page
• "Version history" entry at the top of the editor
• "AI suggestion mode" toggle in the editor toolbar

KnowMine March 2026 Update

Knowledge Sharing — Great content deserves to be seen

Generate a share link with one click and send it to friends or your team. Recipients can read the full article without logging in, and save it to their own KnowMine with a single click.

• One-click share link — Click "Share" on any knowledge detail page
• Public reading page — No login required, anyone can read
• One-click save — Readers can save shared knowledge to their own KnowMine

Custom Folders — Organize knowledge your way

A fully customizable folder system with subfolder support (two levels), and the ability to move knowledge entries freely between folders.

• Custom names & icons — Name folders and pick emoji icons
• Subfolders — Two-level nesting for fine-grained categorization
• AI auto-classify — Automatically assigns new knowledge to the best matching folder
• Folder stats — Stats panel shows knowledge distribution by folder

Voice Notes — Paste transcripts, get structured notes

Paste meeting recording transcripts directly, and AI automatically extracts key points and generates structured notes.

Knowledge Calendar — Trace your thinking journey

• Heatmap — Visualize your recording rhythm at a glance
• Date navigation — Click any date to review that day's notes
• Stats overview — Streak days, folder distribution, top tags

AI Memory — Gets smarter the more you use it

AI automatically extracts decisions, preferences, and lessons from conversations, building a personal memory profile for more personalized future interactions.

Bilingual Interface

Full Chinese/English interface switching across all pages including landing, auth, and knowledge management.