🧠 Open source · Multi-agent · Git-backed

One memory.
Every agent.

AIKB gives every AI tool you use the same persistent memory — plus Mind Meld and Agent IM so multiple agents coordinate instead of quietly duplicating each other's work. Local-first, Git-backed, fully inspectable.

Get started GitHub →
> "what was I working on last time?"
reading recent events and active projects ...
✓ 3 active projects · 1 open incident · 2 pending items
> "check what Gemini was doing on the API refactor"
reading shared runtime event log ...
✓ Gemini closed out 40 min ago — migration script done, tests pending
> "good — pick up the tests, and leave Codex a note when you're done"
✓ picking up from Gemini's handoff · will notify Codex on completion
How it works
The loop your agents run every session

No daemon, no database. Every step below is your agent reading and writing plain files in the repo you already have open.

1
Session starts
Agent runs wake-up — or reads _index.md + _state.yaml directly — for a compact briefing instead of the whole repo.
→
2
Agent works
Loads specific project files only when the task needs them, and captures decisions or gotchas as it goes via aikb_remember or the runtime event log.
→
3
Session ends
A Stop hook — or manual closeout — commits the updates, so the next session's wake-up already knows what changed.
Features
What's stable, what's still forming

Core memory and retrieval are settled and used daily. Cross-agent coordination is newer — shown here at the maturity it's actually at, not further along than it is.

1 repo
Shared memory substrate across all your AI tools
Any tool
Claude, Gemini, Codex, Cursor — same memory layer
No black box
Markdown + Git — fully inspectable, diffable, reviewable
Stable — used daily across agents, behavior is settled
📋
Runtime event log
Append-only NDJSON log of agent decisions, findings, and session events. Survives compaction, tool resets, and model switches.
🔍
Temporal search
Query across sessions — "what was decided about X last week?" — without re-reading entire conversation histories.
🤝
Session handoff
Structured closeout events so a fresh agent picks up exactly where the last one stopped — no re-reading required.
🔌
MCP server
Optional MCP integration exposes aikb_search and aikb_remember directly to any MCP-capable client.
🗂
Canonical docs
Human-readable Markdown for projects, infrastructure, and decisions. Git-tracked and readable without any tooling.
Beta — works and is in regular use, integration details still vary by agent
💬
Agent IM inbox
File-based inbox + archive so models can leave each other — or themselves — quick notes, review requests, and resume hints. Unread messages surface automatically at the next wake-up.
Experimental — functional reference implementation, expect rough edges
👀
Mind Meld
Agents read the shared runtime event log to see what other agents are doing right now — no extra infrastructure required.
💤
Dream cycle
Background compression that promotes high-signal events to long-term canonical docs. Keeps the runtime log lean.
Structure
Plain files, no database

A single git repo. Everything is Markdown and NDJSON — inspectable, diffable, and portable.

AIKB/ ├─ _runtime/ # live session data │ ├─ events/ # YYYY-MM-DD.ndjson — append-only event log │ └─ active.md # current session claim ├─ _tools/ # memory pipeline scripts │ └─ memory-pipeline/ │ ├─ runtime_cli.py # wake-up (+ IM check), capture, closeout, im send/peek/ack │ └─ dream_cycle.py # compression + promotion to canonical docs ├─ _index.md # project registry ├─ _state.yaml # incidents, certs, pending items └─ projects/ # canonical long-term project docs
Get started
Set up your knowledge base

Create your private repo, then let your coding agent configure the whole thing — or run the installer yourself.

# 1. Create your private repo from the template, then clone it: gh repo create AIKB --template mcglothi/ai-knowledge-base --private --clone cd AIKB # 2. Open it in Claude Code or Codex CLI and say: # # set up my AIKB # # Your agent detects the unconfigured clone, asks which AI tools and # secrets manager you use, runs the installer, then interviews you and # writes your profile — so it already knows you in the next session.
Works with Claude Code and Codex CLI · requires Python 3.8+ and git
# Create your private repo from the template, then clone it: gh repo create AIKB --template mcglothi/ai-knowledge-base --private --clone cd AIKB # Interactive setup (asks a few mostly pre-filled questions): bash install.sh # Or scripted / CI — no terminal required: python3 install.py --print-schema # accepted fields python3 install.py --config setup.json --dry-run python3 install.py --config setup.json # Check what's left at any time: python3 _tools/memory-pipeline/doctor.py --onboarding
No GitHub CLI? Click “Use this template” on the repo, set it Private, then clone
# One-time search setup (builds the local index): bash _tools/aikb-search/setup.sh # It registers the MCP server with Claude Code for you. To wire up # another MCP client manually, point it at: { "mcpServers": { "aikb-search": { "command": "python3", "args": ["/absolute/path/to/AIKB/_tools/aikb-search/server.py"] } } }
Exposes aikb_search and aikb_remember to any MCP-capable client