Introduction to Echo
The memory layer under your AI tools: save anything in one sentence, find it when it matters, warm-start every session — on every platform.
Echo is a memory layer that sits under the AI tools you already use. It connects over MCP to Codex, Claude Code, ChatGPT, Gemini, and Cursor, and it does exactly two jobs: everything worth keeping gets saved. And it shows up again the moment a task needs it.
No file to maintain. No folder to organize. No “where did I write that down.” You talk to your agent the way you already do — Echo handles what deserves to outlive the session.
One number tells you whether it is working: your Context SNR. Everything Echo does — save, search, warm-up — exists to raise it.
Save: one sentence, from inside any conversation
Saving is a sentence you say to whatever agent you are already talking to:
Save this to Echo.
Remember this decision.
Save this failed approach so nobody walks that road twice.
What should you save? decisions, constraints, failed paths, repo understanding, your preferences, product language, research conclusions, task state — if it might save the agent one detour in the future, save it.
Echo does not dump the transcript into storage. It extracts structured memories — each one carrying a description, the supporting details, keywords, and its source — so what you saved can be found by meaning later, not by remembering the exact words you used.
No format to think about. No hesitation. The cost of saving is one sentence.
Search: ask for it the way you remember it
Recall is also one sentence:
Search Echo for how we defined the Context Rebuild Loop.
What did we decide about the v21 asset protection?
search memories: warm-up, failed attempts
Echo returns memories with their evidence — the description, the details, where they came from — not a vague summary that sounds right. That discipline is measurable: on LongMemEval, the most widely used long-term memory benchmark, EchoMem reports 95.8%, against roughly 60% for stuffing full history into the window.
You do not have to remember where you said it. Echo does.
Cross-platform: the boundary is your work, not one tool
Most memory features stop at the edge of one product. ChatGPT remembers ChatGPT. Claude Code remembers one repo's CLAUDE.md. Echo draws the boundary differently: around your work itself.
One real task flows through many tools: repo exploration in Codex, architecture reasoning in Claude Code, product language in ChatGPT, frontend drafts in Gemini. Echo converges those traces into one store — and any connected tool can read them back.
In the warm-up piece we called the status quo the Severance protocol: four innies, none of them allowed in the same meeting. Echo is the meeting.
Warm up: start every session with yesterday intact
Before a new task, one more sentence:
Warm up this task.
Echo assembles a launch pack from your memory store, recomposed against the current task's intent: Goal (what you're building right now), Scene (repo map, hot files, recent relevant commits), Decisions (calls already settled), Constraints (boundaries that must hold), and Autopsies (approaches that already failed).
And when a running session grows heavy and drifts toward compaction, don't fight it:
Spin up a shorter, cleaner session.
The live state carries over. The baggage stays behind.
Why this raises your Context SNR
Every memory Echo recalls replaces a re-derivation loop: a grep the agent does not run, a file it does not re-read, a constraint it does not rediscover on turn 40. The signal arrives as a few thousand tokens of memory instead of a few hundred thousand tokens of execution history.
That also makes short sessions affordable. The reason people let sessions run long — and rot — is that starting fresh used to mean losing everything. With memory underneath, a fresh session starts at full context and near-zero noise. Signal stays. Noise stays behind.
Memory is not more context. It is less — only the slices the current task actually needs.
Start today
Save“Save this to Echo.”
Find“Search Echo for the old Context Rebuild Loop.”
Start“Warm up this task.”
Switch“Spin up a shorter, cleaner session.”