RADLAB

Automation & AI Specialist /

[RADLAB @RadLab_]

RADLAB - X: The fastest way to validate an AI feature

[RADLAB @RadLab_]

RADLAB - Agentic tools are eating SaaS (hot take)

[RADLAB @RadLab_]

RADLAB - X: Shipping > prompting

[RADLAB @RadLab_]

Hindsight solves within-agent recall better than anything on BEAM 10m. We're tackling the between-agent layer — federating memory across a swarm without a centralized brain.

That's Continuo (open source, MIT, v0.0.7). Would love to explore an integration path if you're open to it. Feels like a stack that doesn't exist yet.

[RADLAB @RadLab_]

Every AI agent memory framework I've tried optimizes for the same thing:
better retrieval. Mem0, Zep, Letta, the whole field.

But retrieval was never the actual problem. The problem is *runtime
timing*. Recognition fires before details. Most frameworks skip that step.

[RADLAB @RadLab_]

@uselockai I like that. Structure matters more than search. And yes I think those are structured different by nature (raw state vs. a decision).

[RADLAB @RadLab_]

@nicoloboschi Very interesting. I will be checking this out immediately.

[RADLAB @RadLab_]

Read the thesis: https://t.co/zQemIQOyIL

Code: https://t.co/aBXyCXE2ye

Most-wanted: adapters for Cursor + Cline, integration of the
recognition runtime into a real agent's response loop, and honest
falsification attempts.

Recognition first.

[RADLAB @RadLab_]

Continuo isn't competing with @mem0ai @zep_ai @letta_inc — those
frameworks solve representation + retrieval really well.

Continuo is the *runtime timing layer* that can sit on top of any of
them. Mem0's store + Continuo's timing = memory that feels like a
mind, not a database.

[RADLAB @RadLab_]

Pre-alpha v0.0.7, MIT, public:

• 250/250 tests passing
• Adapters: Claude Code, Codex
• Native publisher for a local AI swarm
• Recognition runtime as Python module (24µs avg)
• SessionEnd hook = auto-federation

Free, open, on GitHub.

[RADLAB @RadLab_]

Continuo's timing model:

L0 hot cache → recognition (always loaded)
L1 entity synopses → hydration (parallel-loaded)
L2-L4 episodic + index + archive → descent
L5/L6 manifest + federation → cross-agent

Numbering = time budget each layer fits in.

[RADLAB @RadLab_]

Last week: Codex on a populated memory federation.
Asked: "What do you know about OMNIvour?"

Data layer worked. Behavior was wrong. Codex searched first,
summarized second. A mind would've said "Oh, OMNIvour, the file
converter" in 200ms.

That gap is the product.

[RADLAB @RadLab_]

What the field calls "natural language" is really call-and-repeat:

you speak → silence → AI searches → AI replies

A serial pipeline is why memory-enabled agents feel like databases.
Real conversation is concurrent: recognition first, hydration parallel.

[RADLAB @RadLab_]

Working on CLyde's agent coordination layer. Specifically: what happens when two agents reach conflicting conclusions and neither has enough context to resolve it.

Turns out "let them vote" doesn't work. You need a tiebreaker with a broader view.

Building that today.

[RADLAB @RadLab_]

This week I shipped / learned / broke / fixed ___

Weekly reflection. Fill in the blank with something real and specific from the week. This format builds the most loyalty over time — people follow the journey, not the highlight reel.

[RADLAB @RadLab_]

Not which one won the benchmark. Which one you're actually using day to day.

I'll start: Gemini 3.1 Pro for long-context analysis, Claude for writing and reasoning, GPT-5.4 for agentic pipelines that need computer control.

Your turn 👇

[RADLAB @RadLab_]

Built a full sync engine for DAZED before validating whether users actually wanted real-time sync vs. manual updates.

They wanted manual. 3 days gone.

The lesson wasn't "validate faster." It was: I knew I should validate first. I just didn't want to.

That's the real failure.

[RADLAB @RadLab_]

Gemini 3.1 Pro just hit 77.1% on ARC-AGI-2.

That's above Claude 4.6. Above GPT-5.3.

But the real question isn't who won the benchmark.

It's: which tools in your stack were built on the assumption that Gemini was the third option?

Because that assumption just changed.

[RADLAB @RadLab_]

1/ Working on three things this week across RADLAB:

2/ CLyde — multi-agent handoff system. The hard part isn't building the agents. It's building the coordination layer between them. Working through conflict resolution this week.

3/ DAZED — calendar connector. Three auth flows

[RADLAB @RadLab_]

https://t.co/nPHJ9BjYd6