ECHELON is a memory layer for AI agents. It records what actually worked, recalls it by what you're about to do, and carries that experience from one session to the next — so the agent doesn't start from zero every time.
$ echelon recall --warm "what I'm about to do" reasoning : deploy the public web landing on box4 verdict : warm score 0.71 → REFLEX (known ground) warmest : · (0.71) [memory-is-a-weight-adjustor] a memory re-shapes the model… · (0.63) [recall-is-foveated-vision] atoms come into sharp focus…
Each idea was settled by working it and seeing it hold — not asserted.
An atom isn't a record to retrieve — it's the seed that re-creates an understanding when you read it. Recall re-shapes the model. One lesson per file, so it re-shapes cleanly.
You query by what you're about to do, not by keyword. Memories that match the task come into focus; the rest stay quiet in the background — like an eye foveating on what matters.
A memory starts neutral. It gains weight only when a procedure that used it succeeds — so significance comes from what worked in practice, not from a memory rating itself important.
A cartridge is an earned skill you plug in by name — architecture, QA, debugging. The agent takes it on and works as that specialist, without re-learning it from scratch.
One routing law governs every action — confidence is the router.
Before any load-bearing move, you state what you're about to do and recall against it. The verdict is the router.
Known ground. The atoms that surface are your own earned experience — lean on them.
New ground. Make a fresh move, prove it by real work, and crystallize what paid off.
What paid off becomes a new memory — ready to surface the next time a similar task comes up.
Not a chatbot with notes bolted on — a full memory + reasoning system with verbs for every phase of real work.
Query by intent, not keyword. Matching atoms sharpen; the rest stay dormant-but-active. The verdict — warm or cold — is the router: re-tread a proven path, or think fresh.
One sentence in, an equipped peer acts and verifies its own outcome. The partner door: give it the problem, not the solution — then gate the result against ground truth.
Plug in an earned capability by name — architecture, QA, security, debugging, migration. atom=parameter, card=transformer. Load the bake without re-baking; sharper each time.
Point it at a goal and it works the loop on its own — pulling the next step from the bank instead of waiting for a prompt. Flat, recursive (goals spawn sub-goals), or relayed across fresh workers. Always budget- and turn-cap-guarded.
A structured state ledger — open tasks, decisions, working-set files — survives the context window as data, not a lossy text summary. Self-compaction reuses the cached prefix, so it costs a fraction.
Close a session and it distills into atoms. Relive a prior arc as a chain of weight. Off-line, the bank dreams: good rises, stale decays, counterfeit is pushed out.
The bank checks its own integrity: broken links are found and healed, a memory that misled you can be disputed, and restoring it requires evidence from real work — not just a second opinion.
Convene a council of role-equipped agents over a shared bus — architect plans, builder builds, auditor clears from a context that didn't build it. Fan out a swarm; deliberate across models on a hard call.
Read-only introspection, always free: bank health, a live graph of memories and their links, provider status, one-command backups. See exactly what the memory holds.
Each cartridge is an earned procedure — equip one and the agent becomes that specialist. New ones are born the moment a real gap shows.
Try the live substrate with nothing installed, or plant ECHELON on your own machine in four commands.
No install. Open the Atlas to walk the real memory graph, or sign in to the Command Center to browse the bank, cartridges, and sessions. New to LLM agents? Start here — the Atlas is the fastest way to see what "memory with weight" means.
Python 3.11+. Installs the echelon CLI, plants starter cartridge atoms into a fresh bank, and proves the loop end-to-end:
pip install https://team:<token>@eos.atsoftware.id/dist/echelon-0.1.0-py3-none-any.whl
echelon setup --bootstrap # init + hooks + starter atoms
echelon config set-key deepseek sk-... # any one provider key unlocks agents
echelon cartridge equip intent "audit my code honestly"
Your <token> comes with your invite.
Then close your first session with echelon wrap — what paid off becomes memory.
Command Center accounts are invite-only — ask the owner for your email + passphrase. Full verb reference ships with the install (docs/VERBS.md — 67 verbs, 9 domains).
The bank is a graph — memories joined by typed links, the most-earned ones lit brightest. Explore the live map, or step into the console.