File-based agent memory stores what a coding agent learns in ordinary text files. This page lists six sources on it, oldest first. Each has its own page, which links to the original.
Prompts are code, .json/.md files are state
Mario Zechner argues for treating an LLM as an unreliable computer. The prompt is the program, and files on disk hold the state. The evidence is one real porting job on the Spine runtimes. The time saving is self-reported.
Layers of Memory, Layers of Compression
Compares Letta's layered memory with Anthropic's multi-agent research system and Cognition's single-agent advice. It is included for its frame, because Letta's archive is often a vector database. Kellogg says the text is largely AI-written.
Advanced Context Engineering for Coding Agents
Describes a workflow called frequent intentional compaction. Research, plan and status files carry state between fresh contexts. It covers working context more than long-term memory. The evidence is worked examples, and HumanLayer sells tools in this area.
Claude Skills are awesome, maybe a bigger deal than MCP
Willison describes Anthropic's new Skills feature on the day of its launch. The post explains how skills load only when needed and compares them with MCP. Its evidence is one trial of a published skill and the author's own experience.
Show HN: Recall
A thread of about 85 comments on a local project memory tool. Commenters describe their own practices and question whether such tools are needed. This is the only source here made of personal reports.
Agent memory as a file format
Proposes memoryfield, an open format of Markdown pages with an optional SQLite vector index. It argues that memory should be data rather than a process. The case is design reasoning, from the format's author, with no benchmark results.
I accidentally turned LLM memory into program analysis
The thread discusses a pwning.systems post about Lemmalog, a Datalog agent memory. Commenters describe it as aimed at vulnerability research. They share their own memory setups and raise older symbolic AI ideas. The evidence is anecdote and opinion.
Just-In-Time Agent Memory with Runtime Agentic Research
The paper proposes JAM, which stores full histories as files in a folder tree. A trained Researcher model then searches them for each request. The evidence is benchmark tables, ablations and timing from the authors' own system. Code is released.
Agents Don't Need Memory. They Need Documentation.
Argues that agents should read and maintain documents rather than use memory plugins. It introduces the author's own tool, Operator Memory. Its figures are illustrations, not measurements.
The components of agent memory
Describes memory notes, wikis, injected context and an offline pass called dreaming. It reports evaluations of each. It has the most numbers of the six. They come from Anthropic testing its own models.
Connections to other subjects
These links are this wiki's own reading, not claims the sources make.
- The components of agent memory and A sufficiently detailed spec is code, from Checkable specs for coding agents. Both reach for the same Borges story, about a map as large as its empire. Martin's wiki outgrew its documents. Gonzalez warns a spec can swell toward its code. Each shows a record that has stopped abstracting.
- Layers of Memory, Layers of Compression and Visualizing 100,000 Years of Earth in WebGL, from Images made by rules. Kellogg treats memory as a budget, where detail is spent on what matters most. Stam does that literally, giving 206 of 256 height values to the coastline's range. So what is a memory file's shoreline?
- Agent memory as a file format and What Heraldry and Japanese Mon Can Teach Us About Building Visual-Identity Generators, from Images made by rules. Overmyer lets a state machine run generation while a tree holds the design. Paterson keeps memory as data, apart from any pipeline. Both separate what lasts from the process that produced it.
- The components of agent memory and On the Unreasonable Effectiveness of Property-Based Testing for Validating Formal Specifications, from Checkable specs for coding agents. A wrong spec and a wrong note fail alike, because later work trusts them. Feng and Sergey test specs with counterexamples. Martin tests notes against transcripts. Each checks the layer everything else relies on.
- The components of agent memory and Introducing LACE, from Images made by rules. LACE updates links along with the cells they join. Martin's dreaming similarly rewrites wiki pages and repoints their links. What might memory do if links were state rather than navigation?
- Agents Don't Need Memory. They Need Documentation. and A spec the agent can't run isn't a spec yet, from Checkable specs for coding agents. Liao has the working agent write its own documents. Lewis says anything an agent writes mid-task grades itself generously. Together they raise a question: which memory should an agent be unable to edit?
- Prompts are code, .json/.md files are state and On the foolishness of "natural language programming" (EWD 667), from Checkable specs for coding agents. Zechner calls a natural language prompt a program. Yet the precise work moves into a generated JSON plan, built by a script. Dijkstra's prediction shows up inside the workflow that seems to refute it.
- Prompts are code, .json/.md files are state and Effective harnesses for long-running agents, from Harnesses and agent loops. Zechner and Young's team separately reach for JSON on disk. Each new session starts cold, so progress must live in a file. Zechner's plan comes from a script, and Young's from an agent. Which author of the plan can you trust more?
- Layers of Memory, Layers of Compression and Context engineering, from Harnesses and agent loops. Kellogg reads multi-agent research as an exercise in compression. Loy says agents hand off token sequences, and calls that hand-off an API contract. Then a contract is a statement of what a compression must preserve.
- The components of agent memory and Stevens: a hackable AI assistant using a single SQLite table and a handful of cron jobs, from Databases and storage. Martin calls memory a prediction of what will be needed later. Litt's assistant predicts with a calendar. It sends the coming week's entries plus undated background. A date column may be the cheapest forecast of relevance available.
- I accidentally turned LLM memory into program analysis and Bloom Filters, from Databases and storage. Commenters report models that keep using conclusions already disproved. A standard Bloom filter has a similar difficulty: items cannot be deleted. Counting variants allow it, but give up the promise of no false negatives. Forgetting may be the expensive operation in any memory.
- Layers of Memory, Layers of Compression and The two Christian saints who are secretly the Buddha, from Occult, myth and religion. The Buddha's story reached Christian saints through many retellings, and Bodhisattva became Budhasaf. Kellogg treats memory layers as lossy compression. The legend shows what can survive many hand-offs: plot, not provenance. Historians took centuries to accept the source.
- Agents Don't Need Memory. They Need Documentation. and Adipose tissue retains an epigenetic memory of obesity after weight loss, from Science and medicine. Liao says stale snippets survive because retrieval ranks by similarity, not currency. Hinte and colleagues find fat cells keeping obesity-linked marks after weight loss. Memory can outlast the reason for it. Neither system seems able to mark an entry as outdated.
- Just-In-Time Agent Memory with Runtime Agentic Research and Hymn to Babylon, missing for a millennium, has been discovered, from Occult, myth and religion. Yan and colleagues keep raw sessions because later questions may need discarded details. Babylon's hymn returned because many copies survived, and a platform found 30 more manuscripts. Nobody copying them knew the question. Raw copies let a later reader ask it.
Glossary
- Ahead-of-Time memory
- Memory that is summarized or organized before any request arrives.
- Answer set programming
- A logic approach that can retract conclusions when new facts arrive.
- BM25
- A classic keyword-matching formula for ranking documents.
- Claude Skills
- Anthropic's feature of folders with instructions and scripts that Claude loads when relevant.
- CodeQL
- A tool that turns code into a queryable database for finding bugs.
- Compaction
- Replacing a long session history with a shorter written summary.
- Context window
- The text a model can see at once.
- Cyc
- A decades-old project to hand-encode common-sense knowledge as logic.
- Datalog
- A small declarative logic language that derives new facts from facts and rules.
- Dreaming
- An offline pass that reviews transcripts and notes, then corrects them.
- F1
- A score combining how many right answers were found and how few wrong ones.
- Frontmatter
- A small block of fields at the top of a Markdown file.
- Frontmatter YAML
- A short block of labeled metadata at the top of a Markdown file.
- Graph RAG
- Retrieval for LLMs that searches a graph of linked facts.
- GRPO
- A reinforcement learning method that scores groups of sampled attempts against each other.
- Injected context
- Text loaded into every session at the start.
- Jq
- A command-line tool for reading and editing JSON files.
- Just-In-Time memory
- Memory that is searched and assembled when a request arrives.
- MCP
- Model Context Protocol, a standard for connecting models to outside tools and data.
- Prompt injection
- An attack where text in the model's input tricks it into unwanted actions.
- Provenance
- A record of where a fact came from and how it was derived.
- RAG
- Retrieval-augmented generation: finding stored text and adding it to a prompt.
- Sandbox
- A restricted environment that limits what code running inside it can damage.
- Semantic search
- Finding text by meaning, using numeric embeddings of that text.
- SFT
- Supervised fine-tuning, which trains a model on worked examples.
- Spine runtimes
- Libraries in many languages that play 2D skeletal animations made in Spine.
- Token
- A small chunk of text that a model reads or writes.
- Vector index
- A store of those embeddings, built for fast lookup.