Compare
AI agent memory tools, compared honestly.
Mem0, Zep, Letta, Supermemory: the shortlist people weigh when they search for an AI memory tool or a Mem0 alternative. Here is what each is genuinely good at, what they share, and the one question none of them was built to answer.
The tools people actually compare.
These are strong products. If your problem is agent memory, one of them is very likely the right pick. The benchmark scores you will find elsewhere move month to month, so this stays on what each is for.
Mem0
A memory layer that extracts facts and preferences from conversations and recalls them later, with an open-source core and a hosted option.
Good at: Giving an assistant a durable sense of who you are across chats.
Zep
A temporal knowledge graph: it does not just store facts, it tracks how they change over time, so yesterday's answer and today's stay distinct.
Good at: Agents that need to reason over facts that evolve.
Letta (formerly MemGPT)
Self-editing memory: the agent manages its own memory tiers, deciding what to keep in context and what to page out to storage.
Good at: Long-running autonomous agents that curate their own recall.
Supermemory
Semantic memory at scale, multimodal, tuned for fast retrieval across large amounts of ingested material.
Good at: Search-heavy recall over a big personal or product corpus.
What they have in common.
Underneath the different architectures, they are answering the same question in different ways.
- They attach to an agent or a user, and answer the question can you recall this fact.
- They are measured on recall benchmarks: how much of a long conversation the model can retrieve later.
- They surface information when the model judges it relevant, on retrieval.
That is memory: recall of facts, for an agent. It is real work, and they do it well. It is also not the same thing as keeping a project resumable.
The question none of them answers.
Recalling that you prefer TypeScript is memory. Resuming a three-month project, with everything a new session needs to keep going, is a different job. That job needs:
- The decisions, with their rationale and the alternatives that were ruled out, so settled questions do not silently reopen.
- The work still open right now, which persists until someone actually closes it.
- The documents that carry durable knowledge, versioned with provenance.
- The history of how the project moved, session by session.
- All of it readable from whichever tool, model or person picks the project up next.
Memory helps an agent recall a fact. Continuity helps a project stay resumable.
So which should you pick?
If you need an agent to remember facts, pick one of the tools above. Match it to the shape of your recall: a temporal graph if your facts evolve, self-editing memory for a long-running autonomous agent, a semantic store for search over a large corpus.
If you need a project to stay resumable across sessions, across tools, and across the people and agents who touch it, that is project continuity, and it is what Palimpse is for.
These are not mutually exclusive. A memory tool can make your agent sharper inside a session while Palimpse keeps the project itself resumable around it. One recalls your facts. The other preserves your project.
Definition
Project continuity is the ability to resume a project with its current state, history and rationale intact, regardless of which human, AI, tool or session picks it up next. AI memory helps an agent recall facts. Project continuity keeps the project itself resumable.
AI memory vs project continuity How the MCP connection works
Keep your memory tool. Add the continuity.
Palimpse works alongside whatever memory your agents already use. Connect a first project in a few minutes. Solo is free.