claude-mem

You install the open source engine and it hooks into a supported agent setup. As you work, it captures the session automatically and distils it into structured notes - a decision with its reasoning, a discovery, a fix - each one keeping a pointer back to the session it came from. Recall is progressive: the agent searches for a relevant note first and only reads its details when it needs them, so a past session costs a search rather than a replayed transcript. Memory lives in a local database by default. Adding CMEM Pro moves generation onto a hosted observer, syncs those observations between your connected machines, and opens a private MCP link, which is what lets a different client on a different machine recall the same project history. The project is clear about one limit: recall works over synced notes, and ChatGPT conversations are not themselves automatically captured.

• •Automatic session capture - Records decisions, discoveries and fixes while you work in a supported agent, rather than asking you to write notes
• •Progressive recall - The agent finds a note before reading its contents, so recalling a past session does not mean loading the whole transcript
• •Local first, Apache-2.0 - The engine and its SQLite store run on your own machine, usable with no account and no subscription
• •Cloud sync across machines - CMEM Pro keeps captured project memory available when you move between connected computers
• •Private MCP recall - A private MCP endpoint lets compatible clients, including Cursor, Codex and ChatGPT where custom connectors are supported, read the same project history
• •Session history you can search - Decisions and their sources stay browsable, so you can return to the work behind a change rather than just its outcome

Developers who run long multi-session work with a coding agent and keep paying the cost of re-explaining the same architecture, and anyone who switches between agents or machines often enough that context loss is a daily tax. The local engine is the right starting point and costs nothing, which makes this an unusually low-risk thing to try. The paid tiers are worth it mainly if you work across several machines or want memory generation off your primary agent subscription. Read the billing terms first: the usage ladder and its automatic upgrade behaviour are the part most likely to surprise you, and the included allowances are not on the pricing page. Teams should wait or apply to the pilot, because there is no published team price to evaluate. If you work on one machine, in one agent, on short self-contained tasks, the problem this solves may not be one you have.