v0.1.57 — Your agents stop forgetting
Released 2026-06-20. Multi-platform signed/notarized DMG (Mac Apple Silicon + Intel) + EXE + AppImage. Auto-update from v0.1.56 happens on next launch.
A model is a brain you rent — you and your competitor rent the same one. The memory system is the part you own, and it’s where an agent’s value compounds. v0.1.57 makes memory a first-class, managed substrate for the agents you build in the Foundry: they accumulate knowledge instead of just accumulating logs, and you don’t hand-wire any of it.
Plus two on-ramp wins: the moment you create an agent you land on running it, and the plain-language path needs no API key.
A managed memory substrate
Every agent you build now gets the four memory layers managed for you, on top of CodeNow’s cross-LLM ledger:
- Short-term, budgeted — context is budgeted like RAM (distilled facts first, recent activity next) and assembled cache-friendly, so you ship fast, cheap inference without becoming a KV-cache expert. A
memory.ts/memory.pyadapter is scaffolded into every agent, so a built agent canremember()what it did andrecall()a budgeted memory block on its next run. - Long-term, with a consolidation loop — a background job distills raw episodic activity into durable semantic insights (“here’s what this project reliably wants”) — the engineering analogue of sleep. Runs on CodeNow-served inference; no key required.
- Governed by default — every memory carries provenance; anything an agent writes is wrapped so a poisoned “fact” can’t smuggle instructions into a later run (poisoning defense); and a forgetting sweep retires entries past their TTL — still auditable, never silently lost.
It’s all visible in the Foundry’s Knowledge & Memory step: live stats (episodic / semantic / consolidated / forgotten), recent entries, and one-click Consolidate and Sweep.
Run it now — the default outcome is a working agent
Create an agent and you land straight on running it, not on a static card or a file tree. The artifact you show a colleague is a working agent, minute one.
Zero-setup on-ramp
The plain-language path (“describe it and we’ll build it”) now runs on CodeNow-served inference — no API key to paste, nothing to install. The fastest way from an idea to a working agent.
Upgrade notes
| Change | Who notices | What to expect |
|---|---|---|
| Memory substrate | Anyone building an agent | The Foundry’s Knowledge & Memory step is now a live memory panel; new agents are scaffolded with a memory adapter. Existing agents are unaffected until rebuilt. |
| Run-it-now / zero-BYOK | New builders | After create you land on the run surface; the describe path needs no key. |
Auto-update from v0.1.56 happens on next launch. No migration steps.
What’s next
- Reconstructive retrieval — distilled insights already surface first; graph-structured, multi-hop reconstruction is the next step beyond nearest-neighbor.
- Shared institutional memory — a single governed memory a whole Fleet of agents can inherit, with per-agent private vs shared scoping.
- One-click live deploy — publish a built agent to a callable URL.