CoreSpeed Memory
The memoryevery agent shares.
Say it once, to any agent. Every agent on your team remembers it.
Open to every agent.Closed to everyone else.
Plug in any agent. It's already up to speed.
Any agent.
Claude Code, Codex, Cursor, OpenClaw, Hermes. If it speaks MCP, it remembers.
Private. Until it isn't.
Yours by default. Your team's the moment you share it.
Sealed.
Nobody outside your organization reads a word.
Top-tier accuracy.Same model. Higher score.
Our memory engine on LongMemEval-S, beside each system's own published numbers.
CoreSpeed on LongMemEval-S
94.2with GPT-5.4 answering
- CoreSpeed
- Others, as published
- Our lead
LongMemEval-S accuracyHigher is better
CoreSpeed
GPT-5.4
Mem0
GPT-4o
Zep
GPT-5.4
Supermemory
Gemini 3 Pro
Median search timeLower is better
CoreSpeed
Mem0
Not published
Zep
Supermemory
LongMemEval-S accuracyHigher is better
Median search timeLower is better
- 0 of 500 private memories leaked in testing. The others publish no such test.
- Teammates' changes wait for review. The others publish no such step.
The accuracy axis starts at 80 so the gaps show. A dashed slot means the system publishes no comparable number. How we measured →
Built for a teamthat changes its mind.
Agents propose. People decide. Nothing gets lost.
Review
- Agents propose, with a reason
- You merge. Or don't.
- Agents never merge
- Stale proposals never land
History
- Three versions back, with who wrote them
- Retire the old. Point to the new.
- Undo any archive
- Linked where it matters
Recall
- Pinned memories, from the first message
- By word, by name, by meaning
- Just the passage that matters
- Searching is free
Agents do the remembering.You do the deciding.
They save as they work. They search before they start. You call the shots.
Claude CodeyoursFrom now on, release notes go out on Thursdays.
Saved to team memoryconventionPermissionsno agent can sweet-talk.
Enforced row by row, in Postgres. Nothing an agent says changes that. How isolation works →
| Who | Read | Change | Merge a proposal |
|---|---|---|---|
| You | Yes | Yes | Yes |
| Your agents | Yes | Yes | No |
| A teammate | Yes | Proposes | Admins only |
| Their agents | Yes | Proposes | No |
| Another organization | No | No | No |
How we built it.In the open.
All posts →
Pull requests for shared memory: proposals, archive and versions
A correction next to a stale shared memory does not stop the old one being retrieved. Archive it with a successor, or propose a fix for its author to merge.
Memory isolation with row-level security, not application code
Caller identity comes from auth, each memory carries its scope, and Postgres row-level security enforces the boundary. Nothing in the request body can cross it.
How memory search works: hybrid retrieval, excerpts and result budgets
memory__search fuses exact and relaxed lexical matching, entity aliases and vector similarity into one ranking, then bounds each row and the result.