CrewAI Integration

Give your CrewAI agents persistent, semantically searchable memory that survives across crew runs. Conversations are automatically crystallized into Neural Snapshots — structured knowledge atoms with context, evidence, and connections.

Instead of each crew run starting from zero, agents recall what they learned last time and build on it.

Installation

bash

pip install neuralsnap-crewai

Quick Start

python

from neuralsnap_crewai import NeuralSnapCrewMemory

# Initialize memory
memory = NeuralSnapCrewMemory(
    api_key="ns_your_api_key",
    brain_id="brain_abc123",     # optional — scope to a specific brain
    auto_store=True,             # auto-crystallize significant interactions
    auto_recall=True,            # auto-search for context before tasks
    store_interval=5,            # crystallize every 5 interactions
    recall_limit=5,              # return top 5 context items
)

# Save a crew interaction
await memory.save(
    role="researcher",
    content="PostgreSQL outperforms MySQL 3x for our analytical workload",
)

# Search for relevant context
results = await memory.search("database performance")
for r in results:
    print(f"[{r.score:.2f}] {r.name}: {r.core}")

# Get formatted context for prompt injection
context = await memory.get_context("database selection criteria")
# Returns: "1. PostgreSQL outperforms MySQL 3x for analytical workloads..."

# Force-crystallize buffered interactions
await memory.flush()

Using with CrewAI

Integrate with CrewAI's task system to give agents context from previous runs:

python

from crewai import Agent, Task, Crew
from neuralsnap_crewai import NeuralSnapCrewMemory

# Shared memory across agents
memory = NeuralSnapCrewMemory(api_key="ns_...")

# Before a task runs, get context
context = await memory.get_context("market research findings")

researcher = Agent(
    role="Market Researcher",
    goal="Analyze market trends for Q1 2025",
    backstory=f"Previous research context:\n{context}",
)

task = Task(
    description="Research competitor pricing strategies",
    agent=researcher,
)

crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()

# After the task, save what was learned
await memory.save("researcher", result)
await memory.flush()  # ensure everything is crystallized

Configuration

Property	Type	Description
`api_key`*	string	Your NeuralSnap API key (starts with ns_).
`brain_id`	string	Scope all operations to a specific brain. If omitted, uses the default brain.
`base_url`	string	Override the NeuralSnap API URL. Default: https://api.neuralsnap.ai/v1
`auto_store`	boolean	Automatically crystallize significant interactions. Default: true.
`auto_recall`	boolean	Automatically search for context before tasks. Default: true.
`store_interval`	number	Crystallize every N interactions. Default: 5.
`recall_limit`	number	Max context items returned per search. Default: 5.

Methods

Property	Type	Description
`save(role, content)`	async	Save a conversation turn or crew interaction to the buffer.
`search(query, limit?)`	async → list	Search for relevant memories. Returns MemorySearchResult objects with id, name, core, score.
`get_context(query)`	async → str	Get a formatted context string for prompt injection.
`flush()`	async	Force-crystallize all buffered interactions.
`clear()`	async	Clear the buffer without crystallizing.
`buffered_count`	int	Number of interactions currently buffered.
`total_turns`	int	Total interactions recorded since initialization.

Early access

The CrewAI integration is in early access (v0.1.0). The save, search, and flush methods are fully defined but connect to the NeuralSnap API in v0.2.0. Install now to lock in the API surface.

Fathom

LlamaIndex