Every coding agent starts each session from zero. You re-explain the project, the decision from last week and the reason for it, and the correction you have already made twice. The launches below all attack that problem, and they disagree about almost everything else: where the memory is stored, whether an agent may write to it, and whether a human reads it before it lands.
Memory as plain files on your disk
Renor takes the simplest position it can. What you tell Claude Code, Codex or Cursor is kept in a folder of Markdown files on your own machine, and served back to any of them over MCP. One agent saves a decision, the next one finds it, and the memory is plain files you own. It runs locally with no account and no API key, which is the whole appeal: there is no service to trust and nothing to export later.
Memory that has to be earned
MuscleMemory is procedural memory that learns from what worked and what failed. It watches your agent work, and when a task really succeeds — the tests pass after the last change — it records how the task was done. The next time something similar comes up, it hands the agent a short note instead of a transcript. Writing only on verified success is a filter most memory systems skip, and it keeps the store from filling up with confident failures.
One base for several agents
Busabase goes wider. It is an open-source database and workspace where Claude Code, Codex, Cursor and your own agents keep the records, docs, skills and apps the next task builds on: structured data, durable knowledge, reusable skills, runnable apps, and human review on the writes that matter. It is local-first and self-hostable, so different agents share one base rather than each keeping its own notes.
Context your team can also read
Rinkata puts humans and agents on the same source of truth: goals, specs, decisions and completion evidence, working with Codex, Claude, ChatGPT, Cursor, Grok and Gemini. Structure keeps agents from inventing context, and completed work is added to the knowledge base automatically, so nobody has to nag engineers into documenting. You can upload designs and docs and extract ideas from them with one click.
Memory switched on before you ask
TinyWebUI is a small local-first chat workspace for any OpenAI-compatible model that ships with web research and long-term memory already connected. Chats, documents and tool output all live in one SQLite file on your machine. The point is that there is nothing to configure: memory and search are on by default rather than a plugin you discover later.
Understanding that outlives the run
CommonForge is built for many agents and people changing one project at once, running on Cloudflare Workers and Artifacts, so the project's understanding survives the agent. That framing matters more than it sounds. Most memory tools store what an agent learned; this one is trying to store what the *project* means, independent of who or what worked on it last.
An assistant that grows on its own
CowAgent is an open-source agent harness that plans tasks, runs tools and skills, and grows with memory and knowledge. Its memory is three-tier — context, daily, core — with automatic distillation and hybrid keyword plus vector retrieval, and it installs with one command to run around the clock on a laptop or server. It is the most ambitious of this group, and correspondingly the most to configure.
Memory shared by a group, not a person
Sugabots moves the unit of memory from the individual to the pod. A pod is a shared space for a group of people and their agents where everyone sees the same chats, results and approvals, and agents ask before they act. Agents can even work with each other, such as one checking a trip against the budget another one watches.
Remembering in order to spend less
Wiyomi treats memory as a cost control rather than a feature. It is a developer agent for macOS and Windows whose built-in strategies, smart routing and memory aim at fewer tokens on the same tasks while it codes, researches and writes, and it remembers useful context and preferences as you work. Not repeating yourself is cheaper, not just nicer.
Teams whose memory you can inspect
Guilds is an open-source alternative to closed agent bots, and its pitch is about trust rather than recall. Multi-agent workflows are easy to start and hard to trust, so it shows you the exact prompt, every tool call and every memory change. It runs on your machine with any model. When several agents share a memory, being able to read the diffs is the difference between debugging and guessing.
What they agree on
Across all ten, two things hold. Memory is moving to files and databases you own rather than into a vendor's store, and the write path is becoming a place where humans get a say — review on the writes that matter, approval before an agent acts, or a test that has to pass first. The disagreement is about granularity: a sentence handed to the next agent, or a whole shared base that everybody reads.