Your project history stays on your machine
The core runtime is a local SQLite database and shell hooks. No remote memory account is required for the main product to work.
Why it exists
Long AI coding sessions should compound.
Agents are good at short bursts. Real projects are not short bursts. They contain decisions, weird fixes, important files, release risks, and unfinished tasks that should not vanish every time context compacts or a different agent opens the repo.
Point of view
Memory needs accountability.
Eagle Mem is built around one product belief: recalled context is useful only when the agent can see where it came from, when it was created, and what safety obligation it implies.
Agent surfaces do not blur together
Sessions, summaries, memories, plans, and tasks record their source agent. Claude Code, Codex, and Antigravity get lifecycle capture, while Grok uses the same memory through skills and CLI commands.
Remembering is not enough
Feature verification and guardrails turn past lessons into concrete release checks before push, PR, or publish.
Who it helps
For people who keep reopening the same repo.
Eagle Mem is most useful when an AI coding agent is not just answering a question, but repeatedly working inside the same project over days or weeks.
Use it when the agent should remember architecture, avoid repeating regressions, pick up task state after compaction, and coordinate with another agent without creating a second memory silo.
Principles
Small pieces. Visible behavior.
The product deliberately favors boring tools that users can inspect and recover. If a hook misbehaves, it should be visible. If memory is stale, it should be correctable. If a release is blocked, the user should know why.
No hidden daemon
Hooks run at lifecycle points. The database is a file. The installer can be rerun with `eagle-mem update`, and uninstall removes the hook registrations.
No blind memory injection
Recall is project-scoped, source-labeled, compacted, and tied to summaries, memories, tasks, code index entries, or feature checks.
No folder drift
Monorepos are treated as one workspace when the session belongs to one workspace. Moving between apps or shared folders should not make memory counts drop to zero.
Agents run the workflow
Orchestration commands are designed to be invoked by the active agent. Users should see the plan and results, not babysit lane bookkeeping.
Marketing should match runtime
The website, README, npm page, and CLI output now explain the same system: shared memory, regression safety, RTK token guardrails, and worker lanes.
Start
Install it where the agents work.
Run the install once, then open Claude Code, Codex, Grok, or Antigravity in a project. Eagle Mem will create the project memory layer, refresh linked skills and integrations, and apply patch updates by default when new bug fixes are published.
npm install -g eagle-mem
eagle-mem install
# patch updates apply automatically by default
eagle-mem updates status