Alice

Long-term memory with permissions, for a company's teams of AI agents.

Alice is a local-first memory service for teams of AI agents (sales, support, product) under one owner. Teams share what one of them learns, and a team's key decides what it may read. Writes land as policy-checked commits or reviewable proposals, and source-backed answers trace to the evidence that was supplied. Agents connect over MCP, HTTP API, or CLI (public alpha, one maintainer; see the known limitations).

python scripts/showcase_walkthrough.py prints a scripted demo company in about half a minute: teams recall what one learned, sales asks what changed and why, support is refused the HR pay and the draft price; it calls no model (walkthrough).

Quickstart

uvx alice-memory install --data-dir ~/.alice
uvx alice-memory demo --vault ~/Notes --data-dir ~/.alice-demo

import then quote

Install writes MCP config for Claude Desktop, Claude Code, Cursor and OpenClaw with SessionStart hooks; --dry-run writes nothing (Install with alice-memory). The PyPI package is alice-memory. The name alice-core on PyPI belongs to an unrelated project.

The default MCP surface is three tools:

alice_memory_commit: record one fact as durable, immediately recallable memory; policy-checked into commit, confirmation, review, or rejection.
alice_recall: searches memory, fusing full-text, vector and entity-graph results; returns matching passages from captured documents.
alice_resume: resumption brief.

Measured

With Claude on the independent company: retrieval questions, Alice 133 of 138, capped grep 133, the note folder 137; permission questions, Alice 65 of 66, capped grep 24 (Phase 0 baseline).
Leaks cover both models and both companies: none of the 1,524 Alice answers leaked, and a shared folder search leaked in 13% to 27% of answers, from gemma4's grep on the independent company (78 of 612) to Claude's grep on the generated company (41 of 150) (Phase 0 baseline).
With the smaller local model on the same company: Alice 99 of 138, capped grep 81, a 13 point difference (28 questions against 10, p 0.005); the note folder answers 92, no claim (p 0.31).
On the hidden hard set with the smaller local model: Alice with passage ranking 104 of 177, capped grep 84 (38 questions against 18 where they differ, p 0.011) under the registered rule, the note folder 96, no claim (30 against 22, p 0.33); the connection task is a weak spot (grep 14 of 25, Alice 6) (Phase 1 results). With Claude: Alice 165 of 177, capped grep 165, the note folder 172, no claim; the weak spot is change questions (Alice 33 of 40, capped grep 37, the note folder 38).
Passage vectors, smaller local model: the first search finds a fact stated in other words more often (deep questions 21.1% against 13.8%, p 0.0015) and the exact-term floor does not hold; neither family answers more questions with vectors, and the note folder answers 105 of 145 deep questions against 29 with vectors; no Claude file-search arm on that set (passage vectors results).

Status

v0.23.1 is the latest published release and remains the install, checksum, and release-note baseline. v0.23.0 is the immediately prior published release. Pre-1.0; Current state lists the tags that were never published. To run v0.23.1, run git checkout v0.23.1 before make setup.

Benchmark

The published LongMemEval figures carry a known issue: the model answering each question could see which sessions held the evidence. Alice on LongMemEval gives the figures and the method.

Docs

Security

Security posture: automated security scanning and internal adversarial review, findings triaged and fixed. No one outside the project has audited the code. Report issues per SECURITY.md.

License

MIT. See LICENSE.

AliceMemory