Concept
What Is AI Work Memory?
AI work memory carries sessions, decisions, lessons, project context, and wiki pages across tools and time.
01
Captures decisions, preferences, gotchas, and project knowledge from AI work.
02
Makes memory visible and correctable instead of hiding it inside a model profile.
03
Lets multiple AI tools recall the same durable context through MCP.
01
The short definition
AI work memory is durable context from real work with AI agents, made available when a later session needs it.
That context can include decisions, facts, project constraints, personal preferences, lessons learned, handoffs, wiki pages, and relationships between ideas. The goal is simple: your AI should not rediscover the same knowledge from scratch every session.
02
Memory options for AI coding assistants
Claude Code, Codex, and Cursor start each session without the previous conversation. Most developers combine three layers.
- Instruction files (CLAUDE.md, AGENTS.md, Cursor rules): short standing rules loaded every session, not a store for accumulated decisions.
- Native memory, such as Claude Code auto memory: notes one client keeps for itself on one machine.
- An MCP memory server: decisions, lessons, and project facts any connected client can search on demand.
- Wenlan is that third layer, with source links, review, and Spaces per project.
03
What to capture
Capture something when a future AI session would waste time or make a worse decision without it. Good captures are durable, atomic, specific, and say why the fact matters. Wenlan's memory types are identity, preference, decision, lesson, gotcha, and fact.
- Decisions, and why the chosen path won.
- Gotchas that would cause repeated debugging.
- Corrections, naming the memory they supersede.
- Project constraints that are not obvious from source files.
- Skip raw logs and temporary todos. Capture the durable conclusion, root cause, or command that proved the fix instead.
/capture We chose source-backed pages because summaries need provenance.
/capture Supersedes mem_abc123: Windows setup now uses a Task Scheduler ONLOGON task.
Bad: /capture tests failed
Better: /capture Gotcha: the build fails if Learn relatedSlugs point to missing article slugs; run the slug audit before building.04
Why built-in memory is not enough
Built-in memory is convenient, but it is usually opaque. The assistant decides what matters, stores a compressed version, and may retrieve it later without showing you why.
For real work, people need memory they can inspect, correct, delete, and trace back to source conversations. Bad memory is worse than no memory when it contains stale decisions or wrong assumptions.
- You need to see what the assistant remembers.
- You need provenance for important claims and decisions.
- You need memory to move across tools, not stay trapped in one chat product.
- You need contradictions and duplicates to be managed over time.
05
How Wenlan approaches AI work memory
Wenlan is a local-first, source-backed LLM wiki for AI work in Claude Code, Codex, Cursor, Claude Desktop, and other MCP clients, with experimental web access for supported web AI clients.
Wenlan stores useful context locally, makes memory visible and correctable, writes handoffs, distills source-backed wiki pages, and uses hybrid retrieval that combines vector search, full-text search, and graph context.
06
Source trails and project scope
Provenance is the trail from a remembered fact to its source. Wenlan keeps source memory IDs with page records, exposes source memories through page-source views and APIs, tracks revisions, and writes local git history for readable artifacts. Session logs are useful context but not per-fact provenance, and provenance does not make a memory true; it makes claims inspectable so stale ones can be corrected.
Wenlan is for repeated AI work, not one-off chats. Spaces are project or client buckets for context that should not automatically inform each other, not account permissions or team governance. It is not a life OS, a general workflow suite, or a memory SDK for app backends, and connected AI clients may still send retrieved context to their own model providers.
Make your AI work compound
Wenlan turns decisions, lessons, handoffs, and project context into memory and wiki pages your agents can use later.
FAQ