Agent memory as a file format

TL;DR Highlight

Storing agent memory as Markdown plus a SQLite vector index cut retrieval to two tool calls.

Who Should Read

Backend developers who want long-term agent memory but find pipelines with pgvector and graph databases too heavy.

Core Mechanics

Memory is data, not a pipeline
- Existing memory systems are multi-stage processes that extract information from conversations.
- The author defines memory as a data format and drops the process.
A memoryfield is md files plus SQLite in a zip
- It holds Markdown pages (optional YAML frontmatter) and an optional SQLite vector index.
- The example index uses nomic-embed-text-v1.5.
Agents write prose directly, no chunking
- The agent writes the memory in Markdown at creation time, so no chunking or summarizing is needed.
- Pages are capped at about 8kb (~2000 tokens); add more pages for more detail.
Semantic jump instead of graph walking
- Walking a graph needs N+1 tool calls for depth N, at roughly 2-3 seconds each.
- Semantic search plus parallel reads needs at most 2 tool calls (one search, one read).
Leave access patterns to the model

Community Reactions

Experience of skipping memory entirely
- One commenter said a single poisoned line hurts everything downstream, so they use per-agent temp folders and prune constantly.
- They treat information that can be inferred as noise and keep it out of documents.
Opinion that a good AGENTS.md beats memory
- One commenter found a really good AGENTS.md almost always gives better results than enabling memory.
Pushback on "irrelevant material never surfaces"
- Commenters said past dead ends or wrong conclusions still look semantically relevant and keep surfacing.
- Another said pure vectors can fail keyword lookup, so hybrid search is still useful.
Suggestion to use a git repo
- A commenter proposed a git repo of Markdown pages for incremental writes, commit pinning and diffs instead of regenerating a zip.
- They said to gitignore SQLite as a rebuildable cache.
Experience with a similar setup

How to Apply

Try the sample memoryfield to see the search flow
- Install ollama, uv and npx, run ollama pull nomic-embed-text, then install the CLI and skill.
- Open the author's soapstones.memoryfield.zip and check whether your agent searches then reads in parallel.
Migrate existing agent memory to md pages
- Rewrite fact lists as prose pages under 8kb with title, summary, uuid and date frontmatter.
- Add source URLs to each page so it can be verified later.
Count tool calls in your graph-based memory
- Check logs for how many tool calls one question triggers in your current graph or wiki setup.
- Switch to one search plus parallel reads and compare call count and latency.
Pin and review memories from outsiders
- Open and read any zip shared with you.
- Pin it with sha256sum and connect only trusted sources to your agent.

Try It

copy & run

# Run in a terminal for Claude Code (install steps as in the original)
# Requires: ollama, uv, npx
ollama pull nomic-embed-text
uv tool install git+https://github.com/calpaterson/memoryfield-tool
npx skills add calpaterson/memoryfield-skill -g -y

# After installing, ask Claude Code something like:
# "Using the memoryfield skill, create ./my-memories.memoryfield.zip and save what I learned today about {{topic}} as Markdown pages under 8kb, with source URLs."

# Next: download soapstones.memoryfield.zip, ask how to search Reddit, and count the tool calls

Terminology

memoryfieldA portable memory format that packs Markdown pages and a SQLite vector index into one zip. Think of a notes folder with a search index attached.

Semantic searchFinds documents by meaning using embedding vectors rather than words. It matches even when keywords differ.

Karpathy wikiA knowledge-graph memory of hyperlinked Markdown files that the agent follows page by page. Like clicking through wiki links.

FrontmatterYAML metadata at the top of a Markdown file between --- lines. It works like an info card with title and dates.

Hybrid searchCombines keyword search with vector search. It catches both exact word matches and similar meaning.