Agent memory needs its OpenTelemetry moment

Nov 8, 2026•mthines•View Original

TL;DR Highlight

Agent memory formats are fragmented, so this post proposes an OpenTelemetry-style shared spec.

Who Should Read

A developer frustrated that memory from CLAUDE.md, Cursor rules and memory MCP servers can't be read across agent tools

Core Mechanics

Four memory formats on one machine
- The author lists AGENTS.md/CLAUDE.md, Anthropic's memory tool (/memories), Copilot's repo facts, and a memory MCP server (JSONL knowledge graph).
- None of the four can read the others.
OpenTelemetry agreed on four layers
- A data model, semantic conventions, a protocol (OTLP) and reference implementations were agreed on, while vendors compete on storage, querying and analysis.
- The author says agent memory is where observability was before 2019.
Existing proposals focus on export
- MacPaw Portable Memory (.mem bundle), Open Memory Protocol (April 2026 draft, ZIP+manifest) and Cognee COGX exist.
- They mostly move memory between vendors, while live addressing, many tools sharing one place in a repo, remains open.
Four questions every format answers
- Where is it true, what is its identity (update vs duplicate), what does it say, and when does it stop being true.
- The real fork is who writes it (person, model, extraction service), so a spec should record that.
Example record: frontmatter plus Markdown
- Address, key, id, kind, written_by, lifecycle and attributes go in frontmatter, and the body is Markdown.
- The author says the field names are illustrative, not a proposal.
Spec = conventions, contract, conformance tests
- It proposes a registry of address levels and attribute namespaces, write/read/list/search/archive operations over MCP, and conformance tests.
- Storage, ranking, embeddings and UI stay outside the spec as competitive areas.

Community Reactions

Opinion: standardizing is too early
- Commenters said OTel came after decades of experience and standardizing now would add friction to experimentation.
- Another warned of a premature abstraction like SOAP or SGML.
Counterpoint: LLMs read any format
- Commenters said agents already read memory files in various formats and a simple skill is enough.
- One expected CLAUDE.md pointing to AGENTS.md to perform about the same as native folders.
Parallel session writes should be in the spec
- A commenter said one session seems able to silently overwrite another's memory.
- They cited Anthropic's agent memory research.
Experience: agent memory is harmful
- One commenter said it remembers the wrong things and creates an inane focus, so they disable it on every harness.
- They tell the agent explicitly to record findings in a file when needed.

How to Apply

List where your tools keep memory
- Today, search your repo and home directory for AGENTS.md, CLAUDE.md, .cursor/rules, /memories and MCP memory stores.
- Note scope, author type and expiry for each to spot duplicates.
Add frontmatter to memory files
- Add fields like key, written_by and expires_at to your team's lesson notes.
- Use them to judge trust and prune stale notes.
Consolidate memory write paths
- Have CLAUDE.md reference AGENTS.md and point agents to one write directory.
- Split files by key so parallel sessions don't write the same file.

Try It

copy & run

# Paste into Claude Code or ChatGPT (simplified version of the post's idea: memory record audit)

Below are the agent memory files in our repo.
{{memory file contents}}

Analyze each memory with four questions and output a table.
1) Where is it true (address) 2) Identity (key, duplicates) 3) Content summary 4) When does it stop being true (expiry, evidence it is stale)
Also label the author type (person/agent/service).

Next step: propose merged files that combine duplicates and drop stale ones, with frontmatter (address, key, written_by, expires_at).

Terminology

OpenTelemetryA shared standard for observability data (traces, metrics, logs). Any vendor's tool can exchange data in the same format.

live addressingMany tools reading and writing the same memory location at once and agreeing on its meaning. Like several people sharing one page in a shared doc.

MCPA protocol for agents to call external tools. The post says it standardizes the call but not the record format.

tombstoneA marker left behind to record a deletion. It lets sync systems know something was removed.

Prompt injectionAn attack that hides malicious instructions in text an agent reads. Shared memory widens who can write it.

Related Resources

Agent memory needs its OpenTelemetry moment

Nov 8, 2026•mthines•View Original

TL;DR Highlight

Agent memory formats are fragmented, so this post proposes an OpenTelemetry-style shared spec.

Who Should Read

A developer frustrated that memory from CLAUDE.md, Cursor rules and memory MCP servers can't be read across agent tools

Core Mechanics

Four memory formats on one machine
- The author lists AGENTS.md/CLAUDE.md, Anthropic's memory tool (/memories), Copilot's repo facts, and a memory MCP server (JSONL knowledge graph).
- None of the four can read the others.
OpenTelemetry agreed on four layers
- A data model, semantic conventions, a protocol (OTLP) and reference implementations were agreed on, while vendors compete on storage, querying and analysis.
- The author says agent memory is where observability was before 2019.
Existing proposals focus on export
- MacPaw Portable Memory (.mem bundle), Open Memory Protocol (April 2026 draft, ZIP+manifest) and Cognee COGX exist.
- They mostly move memory between vendors, while live addressing, many tools sharing one place in a repo, remains open.
Four questions every format answers
- Where is it true, what is its identity (update vs duplicate), what does it say, and when does it stop being true.
- The real fork is who writes it (person, model, extraction service), so a spec should record that.
Example record: frontmatter plus Markdown
- Address, key, id, kind, written_by, lifecycle and attributes go in frontmatter, and the body is Markdown.
- The author says the field names are illustrative, not a proposal.
Spec = conventions, contract, conformance tests
- It proposes a registry of address levels and attribute namespaces, write/read/list/search/archive operations over MCP, and conformance tests.
- Storage, ranking, embeddings and UI stay outside the spec as competitive areas.

Community Reactions

Opinion: standardizing is too early
- Commenters said OTel came after decades of experience and standardizing now would add friction to experimentation.
- Another warned of a premature abstraction like SOAP or SGML.
Counterpoint: LLMs read any format
- Commenters said agents already read memory files in various formats and a simple skill is enough.
- One expected CLAUDE.md pointing to AGENTS.md to perform about the same as native folders.
Parallel session writes should be in the spec
- A commenter said one session seems able to silently overwrite another's memory.
- They cited Anthropic's agent memory research.
Experience: agent memory is harmful
- One commenter said it remembers the wrong things and creates an inane focus, so they disable it on every harness.
- They tell the agent explicitly to record findings in a file when needed.

How to Apply

List where your tools keep memory
- Today, search your repo and home directory for AGENTS.md, CLAUDE.md, .cursor/rules, /memories and MCP memory stores.
- Note scope, author type and expiry for each to spot duplicates.
Add frontmatter to memory files
- Add fields like key, written_by and expires_at to your team's lesson notes.
- Use them to judge trust and prune stale notes.
Consolidate memory write paths
- Have CLAUDE.md reference AGENTS.md and point agents to one write directory.
- Split files by key so parallel sessions don't write the same file.

Try It

copy & run

# Paste into Claude Code or ChatGPT (simplified version of the post's idea: memory record audit)

Below are the agent memory files in our repo.
{{memory file contents}}

Analyze each memory with four questions and output a table.
1) Where is it true (address) 2) Identity (key, duplicates) 3) Content summary 4) When does it stop being true (expiry, evidence it is stale)
Also label the author type (person/agent/service).

Next step: propose merged files that combine duplicates and drop stale ones, with frontmatter (address, key, written_by, expires_at).

Terminology

OpenTelemetryA shared standard for observability data (traces, metrics, logs). Any vendor's tool can exchange data in the same format.

live addressingMany tools reading and writing the same memory location at once and agreeing on its meaning. Like several people sharing one page in a shared doc.

MCPA protocol for agents to call external tools. The post says it standardizes the call but not the record format.

tombstoneA marker left behind to record a deletion. It lets sync systems know something was removed.

Prompt injectionAn attack that hides malicious instructions in text an agent reads. Shared memory widens who can write it.

Agent memory needs its OpenTelemetry moment

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Agent memory needs its OpenTelemetry moment

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources