Three weeks ago, I wrote about losing access to Fable 5 after 72 hours. The government pulled it offline before most developers even got a chance to try it. That article hit a nerve — turns out a lot of people felt the same frustration.
Today, Fable 5 is back. And as of this week, it’s available as a model option on ClawBase.
I expected it to be good. I didn’t expect the gap to be this big.
Quick Recap: What Makes Fable Different
SWE-Bench Pro (real software engineering tasks):
- Fable 5: 80.3%
- GPT 5.5: 58.6%
- Opus 4.8: 69.2%
- Gemini 3.1 Pro: 54.2%
On FrontierCode Diamond — the hardest coding benchmark available — Fable 5 scored 29.3%. GPT 5.5 scored 5.7%. More than five times the performance on the tasks that actually matter.
But benchmarks don’t tell the full story. What matters is how it performs inside a persistent agent with real memory. That’s what we just tested.
What Changes When You Run Fable Inside a Persistent Agent
ClawBase has always been model-agnostic. You bring your own API key — Opus, Sonnet, GPT-5.5, Gemini, whatever you prefer. The platform handles the rest: always-on uptime, multi-channel access (Telegram, WhatsApp, Slack, Discord), and a 6-layer memory stack that persists across every session.
Adding Fable as a supported model was straightforward. The results were not.
The memory stack performs differently with Fable. ClawBase agents run on a 6-layer memory architecture: conversation compression, persistent files, semantic search, curated facts, knowledge graphs, and temporal awareness. Every model we support can use this stack. But Fable doesn’t just retrieve from it — it synthesizes across layers in a way other models don’t.
Ask a Sonnet-backed agent “what changed in the auth module since last week?” and it gives you a decent summary from the journal entries. Ask a Fable-backed agent the same question and it cross-references the journal with the knowledge graph, notices a related config change it remembers from two weeks ago, and flags a potential conflict you hadn’t thought of.
Same memory. Same stack. Different intelligence applied to it.
Multi-file refactors go from “mostly works” to “just works.” Fable’s million-token context window plus the persistent project memory from ClawBase means it can hold your entire codebase structure in context while pulling conventions and patterns from memory. The “change this pattern across 47 files” task that other models start hallucinating on by file #30 — Fable handles cleanly.
Context re-loading drops to near zero. Every ClawBase agent already remembers your project, your preferences, your active work. But Fable uses that context more effectively. With other models, I’d sometimes still need to nudge the agent — “remember, we use kebab-case for filenames.” With Fable, it picks up those conventions from memory on its own and applies them consistently.
What the Memory Stack Actually Does
For those who haven’t read the deep dive, here’s the short version:
Layer 1 — Conversation compression. Lossless-Claw keeps a DAG of summaries. Your agent never forgets the beginning of a long session, even past the context window.
Layer 2 — Persistent files + semantic search. Plain markdown files — journals, preferences, project notes. Searchable by meaning via QMD (local embeddings, no data leaves your machine). Human-readable, version-controlled.
Layer 3 — Knowledge graph + temporal awareness. Graphiti tracks how knowledge changes over time. Your agent knows you changed your deployment process last Thursday, and why.
Every model on ClawBase gets the same stack. Fable just uses it better.
The Model-Agnostic Reality
ClawBase isn’t a Fable-only platform. You can run Opus, Sonnet, GPT-5.5, Gemini — swap anytime by changing your API key. The memory persists regardless of which model you’re running. If Fable goes down again (it happened before), you switch models and keep everything.
That said, I’ve been running Fable on my own ClawBase agent for a week now, and I haven’t switched back. The gap on complex, multi-step agent tasks — the ones where memory and reasoning compound — is real. Other models do well on simple tasks. Fable pulls ahead on the hard ones.
What This Costs
ClawBase plans start at $16/mo (Lite) and $33/mo (Pro). You bring your own API key — we don’t mark up model costs. Your Fable usage bills directly to your Anthropic account at their standard rates.
The Pro plan includes the full 6-layer memory stack, multi-channel access, and always-on uptime. No Docker. No server admin. No SSL certificates. Setup takes about 2 minutes.
Try It
ClawBase.to — pick a plan, paste your Anthropic API key, select Fable as your model, and your agent is live. The memory starts building from your first conversation.
If you were already running a ClawBase agent on another model, you can switch to Fable from your dashboard. Your existing memory carries over.
Previously: I Had 72 Hours With the Best AI Model Ever Released. Then the Government Took It Away.
Deep dive on memory: I Tested 33 AI Memory Engines — Here’s What Actually Works
What model are you running your agents on? Curious if others are seeing the same Fable vs. everything-else gap.
답글 남기기