Dev.to AI 🤖 Ai 👁 0 📖 5 min read

There is no agent memory format standard. I counted what my agent remembers instead

There's no agent memory format standard as of 16 November 2026. I found no vendor that ships a coding agent in any memory spec effort, the Agentic AI Foundation lists ten projects and none of them is about memory, and th

There's no agent memory format standard as of 16 November 2026. I found no vendor that ships a coding agent in any memory spec effort, the Agentic AI Foundation lists ten projects and none of them is about memory, and the one standards body group on the subject (a W3C Community Group chartered on 19 June 2026) hasn't published a spec. The open proposals are papers and small repositories, and the three with a repository of their own have 2, 7 and 40 GitHub stars.

The question came up twice in one week, Anthropic published The components of agent memory on 6 November and Mads Thines of LoreKit published Agent memory needs its OpenTelemetry moment the next day.

So I'd stop waiting for a standard, and I wouldn't adopt a memory store to get portability either. Put whatever has to survive a change of harness into AGENTS.md and Agent Skills, and treat each tool's own memory as a cache you can lose. I checked what that cache holds on my own machine.

113 memory records, none of them portable

Claude Code auto memory on my development Mac holds 113 records in 33 project directories (counted with Claude Code 2.1.323). I read the front matter keys of each file and nothing else, and here's what the script printed.

memory records: 113
front matter keys (records carrying each):
  description: 113
  type: 113
  name: 113
  originSessionId: 105
  metadata: 97
  node_type: 94
  modified: 54
fields a transfer would need (records carrying each):
  expiry field: 0
  source or citation field: 0
  confidence field: 0
  writer field (agent, model or tool): 0
  body links another memory with [[name]]: 68

Three things in that output matter if you ever want to move these records.

The only provenance goes stale. originSessionId names a session transcript, but Claude Code deletes transcripts after cleanupPeriodDays and the memory files stay.

There are two layouts and no version field, sixteen records have a flat type key and 97 nest it under metadata. One tool wrote both within nine months, so a converter has to guess which one it's reading.

And the links are local, 68 records point at another memory with [[name]], which resolves only inside that directory.

The types say what I'd lose. Of the 113 records, 57 are project state, 40 are feedback and 16 are reference. A feedback record is a correction I gave the agent, so that's 40 standing instructions that live outside the repo, on one machine, where only one tool looks.

And my own portability mechanism turned out to be a symlink 😄

$ ls -l ~/.codex/AGENTS.md
lrwxr-xr-x  20  ~/.codex/AGENTS.md -> ../.claude/CLAUDE.md

Codex and Claude Code read the same global instructions because I pointed one file at the other, and nothing does that for the 113 records.

What already ports, and what does not

Three layers of the agent stack have a shared format today, and memory isn't one of them.

Layer Shared format Who writes it What a switch costs
Instructions AGENTS.md A person Nothing, a symlink or one setting
Procedures Agent Skills (SKILL.md) A person, sometimes an agent Nothing where the harness reads skills
Tool calls Model Context Protocol A server author Re-registering the server
Enforcement None, hooks differ per harness A person Rewriting the hooks
Memory None An agent or a service The memory

MCP is often named as the answer for memory, but it covers the transport only. In Thines's words, "MCP standardises the call, not the record." A memory store reached over MCP moves with you from harness to harness and then ties you to that store.

GitHub Copilot is the other end. Its memory is the best-specified shape I found, facts carry citations, the agent and model that wrote them and an expiry date (90 days by default). It's also the least portable, the docs say "The SDK does not provide APIs to list, edit, or delete memories."

Why I would not wait

The OpenTelemetry comparison argues for patience more than for a spec. OpenTracing and OpenCensus announced their merger on 28 March 2019 and the stable tracing specification came on 10 February 2021, about 22 months later, and that clock started with two projects that already had users and corporate backers. Agent memory doesn't have such a pair. If two major vendors announced a joint format tomorrow, the same pace would put a stable record in the second half of 2028.

The products are also still moving. Gemini CLI's current docs have no save_memory tool, Copilot's default expiry went from 28 days to 90, and Codex stores memories in a directory named memories_v2 with an undocumented file format.

The best case against this

People don't write things down, and agents do. An agent records the correction at the moment it's made, and of my 113 records 79 include a line explaining why the note exists, which is more than I would've written myself 😄 Long instruction files are also followed less well, so moving memory into AGENTS.md has a price.

I accept both points, but they're reasons to keep agent memory switched on and neither makes it portable. A captured note is durable only once it sits where the next agent loads it.

I'd be wrong if one of Anthropic, GitHub, OpenAI, Cursor or Google joined a memory spec effort, if the Agentic AI Foundation opened a memory project, or if two shipping harnesses began loading each other's memory directory. None of that had happened by 16 November 2026.

What I would do

  • If you run two or more coding agents, make AGENTS.md the real instruction file and point the others at it with a symlink or an import.
  • If your agent writes memory, read the directory once a month, move each standing correction into AGENTS.md or a skill (in a commit a person reviews) and delete the note.
  • If a note is project state, such as a decision or a deadline, let it expire with the tool, it belongs in the issue tracker.
  • If you use Copilot Memory, treat its repository facts as Copilot's working notes. Anything the team relies on goes into the repo.

The full post has the seven spec proposals with their status, the five memory shapes side by side and what each vendor changed in the last twelve months: Agent memory format standard: none exists, so keep what must port in the repo.

Have you opened your agent's memory directory? I'm curious what other people found in theirs.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.