Praxis

Praxis gives AI agents shared organizational memory.

Most organizations have important knowledge spread across CRMs, documents, spreadsheets, analytics tools, support systems, meetings, internal notes, and historical decisions. An agent can be smart and still be blind if it only sees one pasted file, one stale export, or one narrow database.

Praxis turns scattered knowledge, operational data, and history into searchable, source-linked context that agents can inspect, explain, and reuse.

It is designed for teams whose work depends on context across disconnected systems: legal firms, managed service firms, healthcare organizations, security teams, product teams, enterprise marketing teams, agencies, consultants, and internal operations teams.

Praxis Core runs locally today. Praxis Reach and Reach for Agencies extend that model toward live operational systems and multi-client workflows. Praxis is not a hosted service, CRM, warehouse, BI tool, or autonomous agent runtime.

Why Praxis Exists

Claude, ChatGPT, and other agents can reason, write, summarize, and act. But they still need the right business memory.

If the answer depends on customer history, operational data, prior decisions, source documents, account notes, incident records, case files, or campaign results, the agent needs more than a prompt. It needs a way to find the right context, understand where it came from, notice when sources disagree, and keep useful knowledge available for the next task.

Praxis is that context layer.

Example questions Praxis is designed to support:

What Makes Praxis Different

Praxis is not just a folder of documents and not just a vector search demo. It is built around the idea that agents should work from memory they can inspect.

Put simply: Praxis helps an agent know what it knows, where it came from, whether it is still current, and when something looks disputed.

Read more: How Praxis Core Is Different, Retrieval Pipeline, and Conflicts And Dedupe.

How Sources Become Agent Context

flowchart LR
    sources["Sources<br/>docs, data, audio, video, live systems"]
    capture["Capture<br/>source IDs, hashes, metadata"]
    memory["Memory<br/>archive, graph, chunks, embeddings"]
    ranking["Context ranking<br/>trust, freshness, relevance, conflicts"]
    explained["Explained context<br/>source trail, warnings"]
    agent["Agent-ready output<br/>answer, report, workflow"]

    sources --> capture --> memory --> ranking --> explained --> agent

    classDef source fill:#f6f8fa,stroke:#8b949e,color:#24292f;
    classDef core fill:#eef6ff,stroke:#0969da,color:#24292f;
    classDef output fill:#fff8c5,stroke:#9a6700,color:#24292f;

    class sources source;
    class capture,memory,ranking core;
    class explained,agent output;

Same flow in words:

The important part is the ranking step. Praxis does not treat retrieval as "nearest chunks only." It keeps context tied to source history, trust, freshness, graph state, entity hints, and conflict signals so an agent can use the result with more awareness.

See It Work

Install Praxis and run the Core demo:

git clone https://github.com/sparkplug604/praxis.git
cd praxis
python3 -m pip install -e .
praxis demo core

The demo ingests a bundled aggregate excerpt from the official Stack Overflow Developer Survey 2024/2025 archives, creates provisional graph memory, chunks it, embeds it locally, and runs an explained search. The bundled source contains aggregate CSVs plus a source manifest with URLs, dates, licenses, and checksums; it does not commit raw respondent rows.

The search step returns an explained result. Read it as:

The raw CLI output includes numeric scoring fields because downstream tools and tests can use them. The practical takeaway is simpler: Praxis shows why context ranked, where it came from, how fresh it is, how trusted it is, and whether conflicts are present.

If praxis is not recognized after installation, use Python's module form:

python3 -m praxis demo core

Windows PowerShell:

py -m pip install -e .
py -m praxis demo core

More setup help lives in Getting Started. If something fails on Windows, PATH, install, SQLite setup, search, or credentials, see Troubleshooting.

What Works Today

Choose Your Path

How The Pieces Fit

Praxis has three related layers.

The goal is not to copy every system into Praxis. The goal is to let agents work from compact, source-linked context while systems of record stay where they are.

For the deeper architecture, read:

• How Praxis Relates To RAG, Skills, And Live Data
• Praxis With LangChain, LangSmith, And Langflow
• How Knowledge Moves Through Praxis
• Trust, Traceability, And Rollback
• Core Architecture

Repository Map

Praxis keeps source-controlled product files separate from generated local data.

Older checkouts may still have generated data in root-level folders such as db/, kg/, vectors/, research/, reach/, or agency/. Praxis prefers workspace/ now, but keeps temporary legacy fallback. Run this to inspect a safe migration:

praxis migrate-workspace --plan

Praxis CLI

The Praxis core is the Python package and CLI. Use praxis --help and praxis <command> --help for exact options.

Full command reference: docs/cli.md

Docs Map

What Praxis Does Not Do Yet

Praxis is early-stage and intentionally local-first.

It is not a hosted service, not a production vector database, and not a full autonomous agent runtime. It does not automatically promote provisional graph updates into high-trust skills or policies, and it does not ship with private corpora or user-specific connectors.

Praxis Reach is also not a live CRM, ads-platform, analytics replacement, or warehouse. The current Reach module ships with mock and fixture connectors, plus experimental HubSpot, Google Ads, Google Analytics, and BigQuery connector classes. Real connectors should be read-only first, source-linked, permission-scoped, and designed so systems of record stay where they are.