Anyone who's written code with Claude Code or OpenAI Codex knows the frustrating experience: you spent yesterday discussing project architecture with the Agent, open a new session today, and it asks from scratch "what does this project do?"
This is the "goldfish memory" problem of AI coding agents — every conversation starts from a blank slate, context doesn't persist across sessions.
agentmemory targets exactly this pain point. Gained 6,500 stars in a week, now at 9,500. And it didn't go viral on hype — the repo includes full benchmarks, claiming "based on real-world benchmarks." Not many projects on GitHub dare to write that.
What Problem It Solves
The biggest value of agent coding tools is understanding your project context. But the reality is:
- Claude Code starts fresh every time —
.claude/directory's CLAUDE.md only stores static rules - Codex sessions don't cross days — yesterday's learned project structure is forgotten today
- You keep repeating "this project uses PostgreSQL, Prisma for ORM, Vitest for testing..."
agentmemory's approach: give Agents "long-term memory" — cross-session recall of project essentials: directory structure, tech stack, coding conventions, historical decisions.
Technical Architecture
agentmemory offers several memory modes:
Entity Memory — structures key project entities (files, modules, dependency relationships). Not just dumping all file contents into a vector database, but understanding their relationships. Knowing what dependencies exist between src/auth/ and src/db/.
Semantic Search — query memory in natural language: "How did we handle user authentication last time?" agentmemory finds relevant historical context.
Plugin Ecosystem — already supports Claude Code (.claude-plugin), Codex (.codex-plugin), Cursor and other mainstream coding tools. Installation is usually just dropping a plugin config file into your project.
Benchmarks — the benchmark/ directory includes performance tests loading 100,000 memories (p50/p90/p99 latency metrics). This engineering rigor is rare in open-source AI tools.
Deployment Options
Supports one-click deployment to multiple platforms:
# Fly.io / Railway / Render all have templates
# Docker is straightforward too
docker run -d -p 3000:3000 agentmemory/agentmemory
Also provides a Python SDK and REST API for use in non-Node.js projects.
Real-World Effect
From community feedback, the most直观 change after installing agentmemory: Agents no longer start from zero each time. They remember:
- Project tech choices and architectural decisions
- Your preferred coding style (functional vs OOP, type annotation strictness)
- What pitfalls were encountered before and how they were solved
- Module dependency relationships
For a mid-size project, this shrinks new developer onboarding from days to minutes — the Agent already has accumulated knowledge from "predecessors."
A Clear-Eyed View
agentmemory's value is undeniable, but it has boundaries:
Memory isn't everything. It mainly stores structured project info, not every line of code change history. After large refactors, memory may need manual updates.
Privacy considerations. Project context (including potentially business logic) is stored in your deployed instance. Self-deployment is safest but means maintaining extra infrastructure.
Competition with native Agent memory. Claude Code and Codex are both developing their own cross-session memory. agentmemory's moat is its tool-agnostic approach — not tied to any single vendor.
Who It's For
- Mid-to-large project maintainers — complex project structure where Agents need context to work efficiently
- Team shared coding environments — multiple developers sharing one Agent instance, memory transfers between them
- Heavy AI tool users — daily interaction with Claude Code / Codex, don't want to repeatedly explain the project
Not suitable for small personal projects. If your project has just a few files, the Agent scanning everything each time is sufficient — no need for an extra memory layer.
Source: agentmemory GitHub · MIT License