Vexic

Memory your agents can trust.

Vexic is a local-first memory core for long-running AI agents. It stores cleaned conversation history, stages candidate memories for review, and promotes durable facts, each one traceable to the messages it came from.

Status: v0.1.8. The current package is a Python 3.13 core with a SQLite reference service, public contract models, retrieval primitives, and conformance tests. It runs entirely on your machine: it reads and writes a local database and never exfiltrates data.

Local MCP works today. A hosted service is coming. See vexic.dev to learn more and join the waitlist. Hosted surfaces in this repository are internal-alpha adapter code, not a public service contract.

Install

pip install vexic

Or with uv:

uv add vexic

Quick Start

Installing the package puts the vexic command on your PATH. Run the local read-only MCP server against a Vexic database:

# POSIX shells (bash/zsh)
vexic mcp-stdio --db-path ./memory.db --tenant-id local --session-id default

# PowerShell
vexic mcp-stdio --db-path .\memory.db --tenant-id local --session-id default

The server exposes two read-only tools to the agent: recall_conversation_history (this session's transcript) and recall_user_memory (durable facts and preferences). recall_user_memory embeds the query locally; install the optional extra with pip install "vexic[local-embed]" to enable it. See docs/usage.md for MCP client setup, the transcript recorder, and smoke-test examples.

From a source checkout (no install), run the same server through uv:

uv run python scripts/vexic-mcp-stdio.py --db-path ./memory.db --tenant-id local --session-id default

Python Quickstart

Use the library directly from Python: append to a session transcript, then search it back.

import asyncio
from pydantic_ai.messages import ModelRequest, UserPromptPart
from vexic import (
    AppendTranscriptRequest, LocalMemoryService, MemoryCapability, MemoryScope,
    Principal, PrincipalType, RedactionContext, SearchTranscriptRequest, TrustBoundary,
)
from vexic.storage import single_message_adapter

scope = MemoryScope(
    tenant_id="local", session_id="session-1",
    principal=Principal(principal_id="me", principal_type=PrincipalType.HUMAN),
    trust_boundary=TrustBoundary.LOCAL_TRUSTED,
    capabilities={MemoryCapability.WRITE, MemoryCapability.SEARCH},
)

async def main() -> None:
    service = LocalMemoryService(db_path="memory.db", tenant_id="local")
    service.init_schema()
    message = ModelRequest(parts=[UserPromptPart(content="I prefer tabs over spaces.")])
    await service.append_transcript(AppendTranscriptRequest(
        scope=scope,
        messages_json=[single_message_adapter.dump_json(message).decode()],
        redaction=RedactionContext(forbidden_values=()),
    ))
    result = await service.search_transcript(SearchTranscriptRequest(scope=scope, query="tabs"))
    print(result.hits[0].body)

asyncio.run(main())

Output: User: I prefer tabs over spaces. All environment variables the package and its adapters read are listed in docs/configuration.md.

Repository Map

• src/vexic/ - memory contract, local service, storage, retrieval, and hosted adapter code.
• tests/ - executable conformance and reliability coverage.
• docs/usage.md - setup, MCP, recorder, operator tooling, hosted-alpha, and smoke-test examples.
• docs/architecture.md and docs/memory-service-contract.md - architecture and contract references.
• docs/adr/ - accepted architecture decision records.
• AGENTS.md - contributor and maintainer workflow rules for coding agents (repo-root source of truth); docs/ai/README.md indexes the supporting notes.

Vexic Console and the marketing website live in the private PyroDonkey/vexic-website repository (open-core boundary); see ADR 0012's addendum.

Package Boundary

The repository root remains uv-managed; do not add Node package files at the root. Console and website ownership moved out of this repository entirely (see Repository Map above). There is no in-repo Node package surface to isolate anymore.

Contributing

Working from a source checkout? Clone the repository and run the test suite with uv:

uv run pytest

See CONTRIBUTING.md for setup and the branch workflow, SECURITY.md for reporting vulnerabilities, and CODE_OF_CONDUCT.md for community expectations.

License

Apache-2.0. See LICENSE.