agent-memory

A personal, cross-agent knowledge base: concepts captured once are retrievable by AI coding agents, in any project, from a fresh session.

Status

Built — all 12 features proved (see IMPLEMENTATION.md for the generated snapshot and BUILD_LOG.md for how). Developed with KodOS, a methodology-first agentic workflow.

What this is

Brent's coding agents each start every session cold — knowledge learned in one session is lost to every other agent and project. agent-memory is a memory harness: an OKF markdown store as the source of truth, plus a rebuildable lexical index (SQLite FTS5), a vector index (local Ollama embeddings — memory content never leaves the machine), and a derived wikilink/topic graph, read through RRF fusion. Loosely descended from the agent-memory-harness capstone; deliberately minus its dreaming and classifier-routing features.

Vendor scope (DECISION_LOG D1): memory work is restricted to approved vendors — Claude (Anthropic) and Codex (OpenAI). Ambient behavior (unprompted save/search) is claimed and proved on Claude; for codex/agy the v1 claim is installation of the AGENTS.md instruction block only. Antigravity/Gemini-backed agents are excluded from memory work entirely.

This repo holds code only — KB data lives in the selected instance root (default ~/.agent-memory; MEM_KB_ROOT selects an alternate instance, D15), each a local git repo with no remote, a hard confidentiality line, and never pushed.