grpcer/ownmem

What it solves

OwnMem provides a governed, git-native memory system for AI coding agents. It solves the problem of "memory poisoning" and unreliable AI recall by ensuring that project knowledge is stored as reviewable Markdown files within the repository itself, rather than in opaque cloud databases. This allows multiple different AI agents (like Claude Code, Cursor, or Gemini CLI) to share a single, deterministic source of truth that can be version-controlled, audited, and rolled back just like code.

How it works

OwnMem separates the act of writing memory from the act of delivering it to an agent:

1. Git-Native Storage: Knowledge is stored in .ownmem/ as Markdown. Trust receipts and metadata live separately from the text to prevent the AI from simply declaring its own content as trusted.
2. Compilation: Instead of reading raw text during a query, OwnMem compiles memory into immutable, content-addressed snapshots for fast, reproducible recall.
3. Deterministic Retrieval: It uses five local candidate lanes (exact, BM25F, n-gram, fuzzy, and graph) to find relevant information without needing an external LLM or network call by default.
4. Delivery Gates: Before information reaches the agent, it must pass four gates (relevance, epistemic validity, task applicability, and action risk). If evidence is insufficient, the system may quarantine the information or abstain from delivering it.
5. Bounded Evolution: The system can automatically update low-risk metadata (R0), but high-risk changes like new prose or policy updates require human review.

Who it’s for

It is designed for development teams who want their AI agents to have project-specific memory that is transparent, reviewable, and migrates with the code. It is ideal for those who prioritize governance and local-first privacy over maximum recall volume.

Highlights

• Cross-Agent Compatibility: Works across various hosts including Claude Code, Codex, Cursor, and Gemini CLI.
• Evidence-Based Trust: Content cannot authorize itself; independent receipts determine if text is delivered.
• Local-First: Default ranking is local and requires no retrieval API bills or network requests.
• Deterministic Recall: The same query and snapshot always produce the same ranking.
• Safe Rollbacks: Automatic edits carry verified inverse operations to restore previous states without erasing history.