Wenlan vs claude-mem: Explicit vs Automatic Agent Memory

Choose claude-mem when you want agent sessions observed automatically, compressed into searchable history, and surfaced through a progressive disclosure flow with little capture-time effort.

Choose Wenlan when you want durable facts, decisions, handoffs, and pages to be explicit, source-backed, reviewable, and readable outside the agent that captured them.

Do not choose between them on the old assumption that only Wenlan works across agents. Current claude-mem source includes Codex hooks and other integrations; current Wenlan uses one local daemon plus MCP and native plugin paths. The useful decision is automatic session history versus deliberately maintained work knowledge.

claude-mem v13.12.4 is an automatic memory-compression system. Hooks capture prompts and tool activity, a local worker processes observations, and later sessions receive compact context or fetch more through search, timeline, and observation-detail steps.

The maintained source uses SQLite with FTS5, optional Chroma semantic search, a worker UI, citations by observation ID, and progressive disclosure so an agent can inspect an index before loading full history.

Its roots remain in Claude Code, but its current repository also includes Codex hooks, adapters, and integrations for other agents. It is no longer accurate to describe claude-mem as a single-tool store.

Wenlan treats memory as maintained work knowledge. `/capture` records one durable fact, decision, correction, preference, or lesson with provenance; `/handoff` records what changed and what comes next; `/distill` deliberately creates or refreshes source-backed pages.

Recall and storage run through one local daemon. Claude Code and Codex have plugin paths, while Cursor, Claude Desktop, Gemini CLI, VS Code, and other clients can reach the same knowledge through MCP.

The default workflow is explicit, but not frozen in manual mode: optional model-backed passes can enrich captures, connect entities, and propose page revisions. `/brief`, `/curate`, and read-only `/lint` keep those proposals and knowledge-health findings inspectable.

Start with the capture contract, not the client list. If missing a useful event is the main risk, automatic observation is attractive. If preserving only deliberate, attributable knowledge is the main risk, explicit capture and review are easier to audit.

Then inspect the artifact you want to own. claude-mem centers a searchable session timeline and compressed observations. Wenlan centers atomic memories plus maintained Markdown pages, citations, revisions, and local git history.

Automatic capture moves effort away from the moment of work. claude-mem records activity and uses an AI processor to compress it, so users do not have to label every useful event in real time.

The later responsibility is retrieval and trust: decide which extracted observation is current, which is merely historical, and how much detail to load. claude-mem exposes observation IDs, citations, search filters, timeline context, and privacy tags to support that review.

Wenlan front-loads more intent. A user or agent chooses what deserves `/capture`, records a focused `/handoff`, and uses `/distill` only when repeated evidence deserves a maintained page. Proposed revisions and conflicts can then be accepted or dismissed rather than silently replacing the source-backed layer.

claude-mem's hooks generate observations and session summaries automatically. Future sessions receive a compact index and can progressively fetch a timeline or full observation details.

Wenlan's `/handoff` asks for a concise record of what changed, what is blocked, and what comes next. The next session can combine that status with relevant atomic memories and maintained pages.

Both reduce repeated setup. claude-mem preserves a compressed history of what the agent observed; Wenlan preserves the durable knowledge and handoff state that a user or agent chose to keep.

Current claude-mem source includes Codex hooks and adapters. Its pinned installation guide lists Claude Code, Cursor, Windsurf, OpenCode, Codex CLI, Antigravity CLI, and OpenClaw as supported IDEs. The exact capture depth depends on each integration, so inspect the maintained setup path for the client you use.

Wenlan exposes one local source of truth through Claude Code and Codex plugins plus MCP connectors for other clients. Its cross-client promise is shared access to the same atomic memories and pages, not automatic observation of every client by default.

For a mixed-agent workflow, compare what each client can write, what it can only read, and how failures degrade. A logo list is not enough.

Choose claude-mem when automatic observation is the feature, not a compromise: you want session history captured with little interruption, summarized by an AI processor, browsable in a viewer, and retrieved in layers.

Choose Wenlan when the durable unit should be an explicit fact, decision, correction, handoff, or maintained page with review and provenance. It is also the clearer fit when readable Markdown and local git history are part of the ownership contract.

Both projects use Apache-2.0 for the compared open-source core. Wenlan's optional desktop app is maintained separately under AGPL-3.0. Licensing therefore is not the useful differentiator; capture and artifact models are.