Skip to main content
Canonical product answers

FactLane, answered from the product contract.

Nine short answers for people, search engines, and AI hosts. Each answer points to the canonical FactLane documentation that supports it.

These summaries do not replace the docs. Follow the source links for the full contract; the live MCP schema and factlane --help-tools remain authoritative for exact tool fields and enums.

  • v0.1.3 supported named release
  • 5 public MCP tools
  • stdio command-launched transport
  • local SQLite + supported Ollama profile
01

What is FactLane?

FactLane is a free, open-source, local-first MCP memory layer for compatible AI-agent hosts. It stores bounded reusable facts—not conversation transcripts—with scope, provenance, freshness, and a Candidate → Current lifecycle.

02

When is FactLane a good fit?

Use FactLane when preferences, project facts, workflow rules, or scoped tool and environment facts should survive sessions, while old memory must not silently overrule current instructions or verified live state.

Canonical sources
03

When is FactLane not the right tool?

FactLane is not a transcript archive, arbitrary-corpus crawler or search index, backup service, managed remote memory cloud, or automatic permission system. In v0.1.3 it is also not an HTTP, SSE, or Streamable HTTP MCP server.

04

How is FactLane different from broad or built-in memory?

Broad or native memory is useful for rich recall and continuity. FactLane is narrower: it governs a small class of reusable facts with exact scope, provenance, freshness, separate verification, and explicit authority boundaries. It is designed to coexist with broader memory, not replace it.

05

What does Candidate → Current mean?

Candidate means a fact is worth retaining for review; it does not mean trusted current truth. Current requires a separate trusted verification step, and even Current memory must yield to newer authoritative instructions or verified live sources.

Canonical sources
06

What are FactLane's authority boundaries?

Memory is supporting evidence, not execution authority. Current user instructions, current repository or product state, and verified live sources can outrank remembered facts. Seeing a tool does not grant write authority, and an ordinary agent cannot grant itself verifier authority.

Canonical sources
07

What are the five public MCP tools?

The public MCP surface is exactly memory_search, memory_get, memory_store, memory_update, and memory_status. The live MCP schema and factlane --help-tools are authoritative for exact request fields, enums, and signatures.

Canonical sources
08

What environment does v0.1.3 support?

The qualified v0.1.3 profile uses Python 3.11+, linked SQLite 3.42.0+, command-launched local stdio MCP, local SQLite and SQLite-vec storage, supported local Ollama embeddings over loopback, and the documented local POSIX storage and recovery contract. There is no automatic remote embedding fallback.

09

How should an agent start safely?

Connect a compatible MCP host in read-only mode first. The first safe success is: the host launches FactLane, exactly five tools are visible, and memory_status succeeds in an exact scope. Enable Candidate writes only after that connection works and only if they are needed.

Next step

Start from a read-only connection.

If the fit above matches your use case, give the project to your agent or follow the exact v0.1.3 Quick Start.