Coding agents stop losing their memory at the tool boundary: one shared store follows you across agents, projects and machines
Synopsis
ai-memory moves a coding agent's long-term memory out of a single tool and machine and onto a server the user runs, so more than twenty coding harnesses read and write one shared store. Previously Claude Code, Cursor and others kept their own notes, which vanished when you switched tools or handed work to a teammate. The source of truth is a git-backed wiki of ordinary Markdown files, with the database as a rebuildable index, and the default path makes no LLM calls, so capture, search and handoffs need no API key.
Interpretation
A coding agent's long-term memory can now be shared across tools, projects and machines, so when one agent stops mid-task another started in the same directory receives a handoff saying where work stopped, what failed and what remains open. Memory features on each platform previously stayed separate: notes lived on one machine, belonged to one agent, and dropped out of view when tools changed or work passed to a teammate. The integration spans more than twenty harnesses, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Grok, Devin, Kimi and Kiro, which all feed and read the same memory.
The source of truth is a git-backed wiki of ordinary Markdown files that can be grepped, opened in Obsidian, edited by hand or synced with rsync. The database is treated as a derived index that can always be rebuilt from those files, so there is no vector store to maintain and no content locked in a binary blob. Lifecycle hooks record prompts, tool calls and session boundaries, sanitize them at a typed privacy boundary before storage, and the server consolidates them into readable pages; the default path uses zero LLM calls.
Memory decays on a tunable schedule, decays more slowly when a page is opened, searched or reached through a link, and cold episodic notes can be compacted into durable facts such as file paths, error codes and decisions instead of being dropped. Near-duplicates collapse into one and likely contradictions are flagged, all without LLM calls; originals stay in git and the version chain and can be brought back with restore-page. Access-weighted retention is always on because it can only keep memory longer, while the parts that rewrite or drop content, namely compaction, dedup and per-tier curves, stay off by default until the user enables them.
Perspective
The system targets individuals and teams using coding agents, with memory on a server the user runs, whether the same laptop, a homelab box or elsewhere. Multi-user auth, per-person attribution and an audit log are built in and are not a paid tier. The cross-project profile of working habits is on by default for a single user and opt-in and private per person on a shared server.
The material reports no controlled experiment or performance numbers, so the practical effect of cross-harness handoffs has no quantitative basis yet. The parts that rewrite or drop content, namely compaction, dedup and per-tier curves, stay off by default until the user enables them. The optional background pass that uses an LLM to rewrite cold notes into whole pages is cut off in the material, so its full behavior is not described in the visible scope.