Skip to content

crewai-case-study.md

Latest commit

 

History

History
134 lines (89 loc) · 3.75 KB

crewai-case-study.md

File metadata and controls

134 lines (89 loc) · 3.75 KB
Raw

CrewAI case study / proof asset

Updated: July 1, 2026

One concrete ai-knot integration story instead of the full product pitch. CrewAI is the best-prepared follow-up surface: large existing pull, memory already a native concept in its docs, and ai-knot plugs into the exact objects CrewAI users already touch.

Official CrewAI references:

The angle

Not "yet another adapter" — instead:

Keep CrewAI's ergonomics, swap in deterministic long-term memory.

That means:

  • Crew(memory=memory) stays the integration point,
  • Agent(memory=memory.scope("/agent/researcher")) stays the agent-level shape,
  • ai-knot takes over long-term storage and ranked recall underneath.

Not abstract "memory" — the developer does not have to replace CrewAI or move to a hosted memory platform to get persistence.

Fastest proof paths

Zero-network proof

Runs without an API key and without CrewAI installed:

python examples/crewai_surface_demo.py

What it proves:

  • the root memory object exists and is usable now,
  • scoped views behave like per-agent memory slices,
  • recall is deterministic and local.

Real CrewAI wiring

pip install "ai-knot[crewai]"
OPENAI_API_KEY=... python examples/crewai_integration.py

If you want ai-knot itself to extract memories from raw CrewAI task output via an LLM-backed provider, use:

pip install "ai-knot[crewai,openai]"

What it proves:

  • AiKnotCrewAIMemory plugs into Crew(memory=...),
  • agent-scoped views work through memory.scope(...),
  • the integration path is native-feeling, not bolt-on.

What to emphasize

Problem

CrewAI users can orchestrate agents well, but long-term memory still needs a persistence story that isn't full transcript replay or a hosted dependency.

What ai-knot adds

  • persistent facts across sessions,
  • deterministic recall with no LLM on the read path,
  • self-hosted storage (SQLite / PostgreSQL / YAML),
  • scoped memory views that map cleanly onto agent roles,
  • optional ai-knot extraction when you want CrewAI task output distilled into facts.

What not to claim

  • Don't say it replaces CrewAI.
  • Don't say it is better because it is "more autonomous."
  • Don't oversell semantic richness; the wedge is reproducibility and storage control.

Copy blocks

GitHub discussion / follow-up comment

A concrete surface that is ready today: CrewAI.

ai-knot now plugs into Crew(memory=...) and agent-scoped memory.scope(...) views, so you can keep CrewAI's runtime ergonomics and add deterministic long-term memory underneath.

Fastest proof:

  • python examples/crewai_surface_demo.py for the zero-network memory surface
  • examples/crewai_integration.py for the real Crew wiring path

X / LinkedIn

CrewAI users: ai-knot now plugs into the native memory surface.

Crew(memory=...) Agent(memory=memory.scope(...))

Deterministic long-term memory, self-hosted storage, no LLM on the retrieval path. Zero-network proof: python examples/crewai_surface_demo.py Full wiring: examples/crewai_integration.py

https://github.com/alsoleg89/ai-knot

Reddit comment or reply

If you want one concrete thing to try instead of the whole repo, start with the CrewAI path. The adapter now fits the native Crew(memory=...) / scoped-agent shape, and there is a zero-network demo in examples/crewai_surface_demo.py before you even wire a real model.

Recommended CTA

Lead with one of these, in order:

  1. python examples/crewai_surface_demo.py
  2. examples/crewai_integration.py
  3. docs/integrations.md

Don't send people to the whitepaper first. Send them to the shortest proof.