Lance Martin's Claude Diary

Table of content

Lance Martin is a software engineer at LangChain who created Claude Diary, a plugin that gives Claude Code the ability to learn from experience and update its own memory. The idea: humans refine their skills through reflection, so agents should too.

Background

PhD from Stanford in computational biology
5+ years leading perception teams at UberATG, IKE, and Nuro (self-driving vehicles)
Early hire at LangChain focused on open source and applied ML
Created LangChain Academy (35+ lessons on agents) and RAG From Scratch (450k+ views on FreeCodeCamp)

GitHub | Twitter | Blog

The Problem with Agent Memory

Most agents are stateless across sessions. They forget what they learned, repeat the same mistakes, and never adapt to your preferences. Martin noticed this gap after extensive work on context engineering.

His solution draws from the CoALA paper (Sumers et al. 2023), which proposes three memory types:

Memory Type	Purpose	Example
Procedural	Instructions that steer behavior	CLAUDE.md rules
Episodic	Examples of desired behavior	Past session logs
Semantic	Task-relevant facts	Project-specific context

Claude Diary focuses on procedural memory, turning session experiences into persistent rules.

How Claude Diary Works

The plugin adds two slash commands:

/diary captures what happened in a session:

What was accomplished
Design decisions made
Challenges encountered
User preferences observed
PR feedback received

Entries save to ~/.claude/memory/diary/YYYY-MM-DD-session-N.md.

/reflect analyzes diary entries for patterns:

Identifies violations of existing CLAUDE.md rules
Detects recurring mistakes
Synthesizes insights into one-line bullets
Proposes updates to ~/.claude/CLAUDE.md

The reflection process tracks processed entries in processed.log to avoid duplicate analysis.

# Manual diary entry
/diary

# Automatic via PreCompact hook
# Triggers before context compaction

What It Learns

After a month of use, Martin’s CLAUDE.md accumulated rules about:

Category	Example Rules
Git workflow	Use atomic commits, specific branch naming
Testing	Run targeted tests first for quick feedback
Code patterns	Prefer composition over inheritance
Token efficiency	Compress verbose explanations
PR feedback	Address reviewer comments systematically

Each rule is a single imperative line. The system strengthens weak rules when they’re violated and adds new rules when patterns emerge.

Architecture

The system follows the generative agents framework:

Observations - Raw session conversations recorded as diary entries
Reflection - Pattern identification across multiple entries
Retrieval - Rules loaded into future sessions via CLAUDE.md

Martin kept reflection manual (not fully automated) so users can review proposed changes before committing them to CLAUDE.md.

Context Engineering Background

Claude Diary emerged from Martin’s broader work on context engineering, which he defines as populating an LLM’s context window with the right information at each step.

Four strategies from his framework:

Strategy	Method
Write context	Save scratchpads and memories for later
Select context	Retrieve relevant information via RAG
Compress context	Summarize long trajectories
Isolate context	Split work across sub-agents

Claude Diary implements the “write context” strategy for procedural memory specifically.

Key Takeaways

Principle	Implementation
Agents should learn	Build feedback loops from sessions to instructions
Memory needs structure	Separate procedural, episodic, and semantic
Reflection beats logging	Transform experiences into actionable rules
Manual review matters	Let humans approve memory updates

Links

Next: Harper Reed’s LLM Codegen