Your AI Agent Forgets Everything. Here's the Fix.

SUMMARY

Jaymin West demonstrates Mulch, an open-source tool that equips AI coding agents with persistent memory across sessions by storing growing expertise in version-controlled repositories.

STATEMENTS

Coding agents often frustrate users by forgetting everything between sessions, lacking continuity in their work.
Mulch is a straightforward tool designed to enable AI agents to build and retain expertise within a project's codebase.
Mulch initializes a version-controlled directory in the Git repository, ensuring that accumulated expertise is accessible to all agents working on the project.
The tool adds hooks for session starts and pre-prompts, similar to the Beads system, to automate memory integration.
Agents use Mulch to create and expand domain-specific expertise files, which evolve over time without unnecessary repetition due to built-in pruning.
Mulch updates configuration files like claude.md or agents.md to instruct agents on using it before starting and after completing tasks.
Through consistent use, agents develop deeper understanding of codebase patterns, conventions, failures, and decisions.
The mulch prime command loads comprehensive context, including project structure, preferences, and domain knowledge, at the start of sessions.
Mulch supports targeted priming, such as for specific domains like blogs, to output relevant expertise only.
As an open-source project on npm and GitHub, Mulch is free and encourages community contributions to identify and fix issues.

IDEAS

Version-controlling AI-generated expertise turns ephemeral agent knowledge into a permanent, collaborative asset within a codebase.
Automating hooks for session starts and pre-prompts seamlessly injects historical context, mimicking human long-term memory for machines.
Pruning mechanisms in Mulch prevent knowledge bloat, ensuring expertise files remain focused and efficient as they grow.
Domain-specific files allow agents to specialize in project areas, fostering targeted intelligence rather than generic recall.
Integrating Mulch instructions directly into agent config files creates a self-reinforcing loop where agents naturally adopt better memory practices.
Agents tracking failures and decisions via Mulch enables iterative improvement, reducing repeated errors across sessions.
The ability to prime only relevant expertise, like blog-specific knowledge, optimizes context without overwhelming the agent's processing.
Mulch's inspiration from Beads highlights how simple, composable tools can layer to enhance AI persistence dramatically.
Building a community around agentic engineering democratizes tools like Mulch, accelerating collective innovation in AI-assisted coding.
Open-sourcing memory tools like Mulch invites real-world testing, turning user-reported issues into rapid evolutions for better reliability.

INSIGHTS

Persistent, version-controlled memory for AI agents transforms isolated sessions into a cumulative learning ecosystem, mirroring human expertise accumulation in teams.
By automating context injection and pruning, tools like Mulch bridge the gap between stateless AI models and stateful human cognition, boosting productivity in software development.
Tracking not just successes but also failures and decisions in expertise files enables AI to evolve proactively, fostering resilience and adaptability in complex projects.
Domain-targeted priming refines AI focus, revealing how selective memory recall can enhance efficiency far beyond broad knowledge dumps.
Open-source communities around such tools accelerate AI engineering by crowdsourcing refinements, underscoring technology's role in collective human flourishing through shared innovation.

QUOTES

"One of the most frustrating parts of using coding agents is that they don't remember anything between sessions."
"Mulch is a very simple tool for your agents to use."
"Expertise um grows and lives in your git repository. So any agent working on this will have access to the mulch expertise that generates over time."
"Over time, what this means is your agents are going to get smarter about your codebase."
"Please go check out mulch. Please go, you know, mess around with it, find the issues."

HABITS

Instruct agents to run the mulch init command at the start of any new project to establish persistent memory foundations.
Consistently execute mulch prime before sessions to load full project context, ensuring informed decision-making from the outset.
After completing tasks, agents should use mulch to record learnings, patterns, and decisions, building a habit of reflective documentation.
Regularly review and prune expertise files with mulch commands to maintain concise, non-redundant knowledge bases.
Integrate mulch setup into agent configurations early, creating a routine where memory tools are non-negotiable for ongoing work.

FACTS

Mulch is heavily inspired by Steve Yegge's Beads project, a GitHub repository focused on similar agent memory concepts.
The tool is published on npm, allowing easy installation via standard package managers for Node.js environments.
Mulch creates domain-based expertise files using simple patterns, starting with three initial domains upon setup.
It adds sections to files like claude.md for Claude AI or agents.md for other systems, embedding usage instructions automatically.
The project includes a free community space called Prompt to Prod for discussions on agentic engineering.

REFERENCES

Mulch Repo: https://github.com/jayminwest/mulch
Steve Yegge's Beads: https://github.com/steveyegge/beads
Prompt to Prod community: https://www.skool.com/prompt-to-prod-...
Consulting services: https://consulting.jayminwest.com/
Claude Code environment for AI-assisted coding
Git repository for version control of expertise
npm for installing Mulch
claude.md and agents.md configuration files

HOW TO APPLY

Begin by navigating to your project repository in an AI coding environment like Claude Code, ensuring no prior Mulch setup exists.
Instruct your agent to run the "mulch init" command, which creates a version-controlled Mulch directory to store emerging expertise.
Follow with "mulch setup claude" to install hooks for session starts and pre-prompts, integrating memory loading automatically into workflows.
Have the agent execute "mulch prime" at session onset to ingest full context on project structure, preferences, and domain knowledge for informed actions.
After task completion, direct the agent to run "mulch" again to update expertise files with new patterns, conventions, failures, or decisions, then commit changes to Git.

ONE-SENTENCE TAKEAWAY

Mulch empowers AI coding agents with persistent, version-controlled memory to build cumulative expertise and enhance codebase understanding over time.

RECOMMENDATIONS

Install Mulch via npm in your next project to immediately address agent forgetfulness and foster long-term intelligence.
Pair Mulch with Beads for complementary memory layers, amplifying persistence in multi-tool AI workflows.
Join the Prompt to Prod community to collaborate on agentic engineering, sharing Mulch customizations and insights.
Use domain-specific priming like "mulch prime blog" to tailor context, improving agent focus on niche project areas.
Experiment with Mulch in open-source repos to identify issues collaboratively, refining it for broader AI adoption.

MEMO

In the fast-evolving world of AI-assisted coding, one persistent headache has been the amnesia of agents like those powered by Claude or OpenAI: they excel in bursts of brilliance but forget everything upon session's end. Enter Mulch, a deceptively simple open-source tool crafted by developer Jaymin West to instill lasting memory. By embedding expertise directly into a project's Git repository, Mulch ensures that knowledge—patterns, conventions, failures, and decisions—accumulates across sessions, accessible to any agent joining the fray. West demonstrates its setup in a bare Claude Code repository, where a single "mulch init" command spins up a dedicated directory, followed by hooks that prime the AI with context before prompts and log learnings afterward.

What sets Mulch apart is its elegance in growth and restraint. Inspired by Steve Yegge's Beads framework, it generates domain-specific files that evolve without redundancy, thanks to intelligent pruning. Over time, agents "get smarter" about the codebase, grasping nuances that once required laborious re-explanation. West highlights targeted commands like "mulch prime blog," which surfaces only relevant expertise, avoiding the overload of irrelevant data. Freely available on npm and GitHub, Mulch invites tinkering—West urges viewers to "mess around with it, find the issues"—positioning it as a communal stepping stone in agentic AI.

As AI tools redefine software development, Mulch underscores a pivotal shift: from disposable interactions to enduring collaboration. West's free Prompt to Prod community fosters this ethos, blending consulting with open dialogue on AI's future. For developers weary of starting from scratch, Mulch isn't just a fix—it's a foundation for agents that learn like teams, promising a more intuitive partnership between human coders and their silicon counterparts.