Skip to content
Wenlan

Privacy

Local-First AI Work Memory: Keep Context on Your Machine

Local-first AI work memory keeps sensitive project knowledge, decisions, and preferences under your control while still making them useful to assistants.

Qi-Xuan LuUpdated 5 min read

01

Your memory database stays on your machine by default.

02

Base retrieval runs locally; optional enrichment can use an on-device model, a local endpoint, or a cloud model you choose.

03

Every memory remains visible, correctable, and traceable.

01

What local-first means for AI work memory

Local-first AI work memory means the durable context your assistants rely on is owned and stored primarily on your device. Cloud services may still be useful in some workflows, but they are not the default source of truth.

For memory, that distinction matters. The data can include client names, strategy decisions, personal preferences, private codebase details, and the accumulated reasoning behind your work.

02

Why memory is more sensitive than prompts

A single prompt may be sensitive. A memory layer is sensitive in a different way because it accumulates. Over time it becomes a compact map of what you care about, what you are building, where you got stuck, and what decisions you made.

That makes visibility and control non-negotiable. You should be able to inspect, correct, export, and delete what your AI remembers.

03

The tradeoff

Cloud memory can be easier to access across devices. Local-first memory gives stronger ownership, simpler privacy boundaries, and better fit for work that cannot casually leave your machine.

Wenlan chooses local-first because the memory layer should be something you trust, not another opaque profile maintained by a platform.

Local-first is not offline-only. A connected AI client sends the context it retrieves to its own model provider, a cloud model you configure for enrichment receives the content it processes, and experimental web access passes queries and results through a relay.

04

How Wenlan keeps memory useful

Local-first does not mean inert. Wenlan combines vector search, full-text search, and a knowledge graph so assistants can retrieve the right work context without replaying everything.

It also makes memory inspectable. You can see what was learned, trace it back to source conversations, and correct it when your understanding changes.

05

Readable artifacts plus a local store

Wenlan keeps Memories and graph data in a daemon-owned local libSQL store that powers fast semantic and full-text recall, and projects Pages with citations and revisions, session logs, and project status as readable Markdown under ~/.wenlan. Markdown alone would not give agents fast retrieval, and a database alone would be opaque to you.

Those readable artifacts are also committed to a local git repository at ~/.wenlan/.git, so changes show up as ordinary diffs and can be recovered. Git versions the artifacts; the daemon still owns the database and indexes.

  • git -C ~/.wenlan log --oneline shows the timeline of artifact changes; git -C ~/.wenlan diff shows readable changes.
  • Back up ~/.wenlan, including .git, with the rest of your local data.
  • Use Wenlan's correction, review, and forget flows for memory changes instead of editing daemon-owned state behind its back.
Inspect local artifacts
ls ~/.wenlan/pages ~/.wenlan/sessions
git -C ~/.wenlan log --oneline -5

Keep your context where your work lives

Wenlan gives AI tools useful memory without making your accumulated work context cloud-first by default.

FAQ

Does local-first mean no AI model can use the memory?+
No. Local-first means the memory layer is owned locally. MCP-compatible AI tools can still access relevant context through the local Wenlan daemon.
Is Wenlan fully self-hosted?+
Wenlan is local-first on macOS, Linux, and Windows. The daemon and database run locally, and optional integrations may depend on the AI tools you connect.
Is Wenlan using git as the database?+
No. The daemon owns the database and indexes. Git versions readable artifacts such as pages, session logs, and project status Markdown.

Share this guide