e-xo
the problem
Every session starts from zero.
You've explained your architecture three times this week.
The bug you fixed last month? You'll fix it again.
The convention you decided on? Gone.
Your AI has the memory of a goldfish.
the fix
What if your AI could remember?
Not chat history. Not vector search over old logs. Actual memory. It strengthens with use, fades without it, and reorganizes itself while you sleep.
e-xo is a knowledge graph that maintains itself. It builds itself from your work, links what you learn across projects and sessions, and gives your agents the right context at the right moment.
Your agents' second brain.
the memory model
This is what your knowledge looks like.
Every node behind you is a memory. Every line is a connection your XO made. Grab one and watch the others follow. The unused ones quietly dim.
Use it or lose it
Memories decay a little every day unless they earn their keep. Preferences and procedures fade slowly; stray notes fade fast.
Only real use counts
Retrieval isn't reinforcement. Agents flag which memories actually helped. Those get stronger; noise gets demoted.
Connections form themselves
Semantic auto-linking wires related knowledge together. Structure emerges from behavior, not from filing.
Typed, with provenance
Episodic, semantic, procedural. Every node carries confidence and origin, and retires cleanly when superseded.
nightly consolidation
At night, it dreams.
Days produce scraps: diary entries, session logs, half-decisions. Every night, e-xo clusters them by theme and consolidates them into durable knowledge. What you know, not just what happened.
And the night is transactional: it measures its own retrieval quality before and after, and rolls the graph back if dreaming made things worse.
consolidation with a rollback plan.
measured, not assumed
Numbers, not vibes.
Retrieval quality is a tracked number here. e-xo ships its own open benchmark harness, runs published memory benchmarks end-to-end, and posts every result, including the unflattering ones.
Today's numbers, republished as they move. Production metrics from the live graph's golden-set evals; LoCoMo accuracy from exo-bench on the full 1,540-question split.
local-first
Actually yours.
Your graph is plain JSON on your machine. Inspect it with cat. Back it up with cp. No cloud in the loop.
~/.exo/ on disk
Every memory is a readable file. Decay, links, provenance: all inspectable, all greppable, all yours.
Or your team's
A lightweight team server gives 2–20 people one shared graph, with join tokens, author attribution, and personal vs team scopes.
install
Beta One command.
curl -fsSL https://install.e-xo.ai | sh
macOS & Linux · arm64 + x86_64 · standalone binary, no Python needed · keeps itself current with exo update
getting started
What happens next
The installer hands off to a setup wizard. Five minutes later your agents remember.
Finds projects in ~/dev, ~/code, ~/projects and imports git commit history as dated diary entries.
Converts commits into memory nodes, linked by date and repo. Your agent gets context going back months.
Detects your coding agents and installs e-xo as an MCP server. No manual config needed.
Installs embeddings so your agent finds the right memories, not just keyword matches.
Opens your knowledge graph at localhost:7797. Browse nodes, connections, and your work diary.
MCP + hooks
MCP server
MCP server
MCP server
MCP server
MCP server
One graph, 19 MCP tools, every agent on the same page.