Memory that keeps its receipts
Extracted facts answered 84.4% of LongMemEval-S. Linking each fact back to the conversation it came from raised that to 89.6%, for 20% more storage.
Tagged
engineering
← All postsExactly once: how an approved agent action runs with frozen arguments
A held call is frozen, the agent waits on one id, and approval hands the call to that waiting request once. Re-sending the call makes a second card.
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.
HTTP 200 with isError true: the most common MCP integration bug
A CoreSpeed tool call can fail inside an HTTP 200. How the two error layers differ, which codes to retry, and which ones need a person instead.
Linking memories: a small graph that keeps decisions connected
Memory links are directed edges with a kind and optional weight. They are idempotent to write, visible only when both ends are, and capped at 1,000 per memory.
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.
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.
What is an audit trail for AI agents?
An audit trail records what each agent did, as whom, and how it ended. CoreSpeed keeps one per org: action, actor, outcome, code and cost, never the prompt.
What is an MCP gateway, and why do agents need one?
An MCP gateway is one remote MCP server that fronts many tools and app accounts behind one sign-in. What that changes for config, credentials, spend and audit.
What is Harness-as-a-Service (HaaS)?
Harness-as-a-Service is a hosted harness any agent signs into. Tools, auth, memory, budgets and an audit trail come as one service, and the agent you run stays yours.
Why reads are never gated: designing an approval gate people keep on
A gate that asks about everything gets switched off. Smart Approval judges writes only, lets reads through, and never lets a failed evaluation run a write.
Why tools/list is yours, and not a catalog
tools/list on CoreSpeed returns exactly what this sign-in can call. Absence is a decision, presence is neither health nor permission, so never hard-code it.
Why your agent should never hold an OAuth token
An agent that holds a token can leak it through a prompt, a log or a transcript. Credential custody lets a platform act with the right account instead.
How we took LongMemEval from 80.0 to 94.2 without touching retrieval
LongMemEval-S climbed from 80.0 to 94.2 while retrieval recall at 10 stayed at 0.99. The gains came from evidence budget, reader generation, and instructions.