Give Claude Code, Codex and Cursor a memory of your repos
AI coding agents are powerful, but they forget how your team already solved repeated problems. Moresq Corpus turns those prior solutions into a searchable, governed memory layer for terminal agents and MCP clients.
The problem: context loss
A coding agent often regenerates authentication helpers, webhook handlers, API clients, analytics events and deployment scripts even when the repo already contains the right pattern. That wastes tokens, increases review time and creates inconsistency between projects.
Repo memory changes the workflow. The agent searches first, retrieves a short verified pattern, then writes only the project-specific adaptation. The source of truth stays in code and proof metadata, not in a long model answer.
Retrieval-first agent loop
- The terminal agent calls `search_corpus` before writing commodity code.
- Moresq Corpus returns compact matches with provenance and quality labels.
- The agent retrieves only the snippet or runbook that should be applied.
- The snippet is applied by tooling, while the model handles names, paths and glue.
- Missed searches become new snippet candidates and evaluation material.
Why this matters for AI agents
Agents get stronger when they can reuse proven local decisions. Moresq Corpus is built for that compounding effect: search hits, retrieval receipts, rejected snippets, abuse signals and apply logs all become product evidence.
The market need is direct: developers want faster agent runs without sacrificing control. Corpus gives agents memory while keeping the human review surface inspectable.
What good repo memory contains
Useful memory is more than a vector match or a copied code block. It names the problem the pattern solves, the assumptions it makes, the interfaces it expects and the tests that protect it. A future agent should be able to decide whether a match is safe before it retrieves the full implementation.
Moresq Corpus keeps these signals beside the artifact: language, provenance, license, scan status, examples, review state and revocation information. That makes memory portable between Codex, Claude Code, Cursor and MCP clients without pretending that one repository convention fits every project.
What memory should not do
- It should not silently override local security, privacy or architecture decisions.
- It should not turn an old snippet into a current dependency without a review date.
- It should not hide uncertainty behind a confident generated answer.
- It should make reuse easier to inspect, test, adapt and revoke.
Frequently asked questions
What is repo memory for an AI coding agent?
Repo memory preserves durable constraints, decisions, reviewed implementation patterns and their evidence so an agent can retrieve the right context in a later session.
Does repo memory replace AGENTS.md?
No. Keep stable, always-on rules in repository instructions. Retrieve detailed decisions and implementation cards only when the current task needs them.
Can the same memory work with Codex, Claude Code and Cursor?
Yes when the underlying records use portable metadata and are exposed through an HTTP API or MCP-compatible adapter rather than tied to one chat interface.