memlink

Pandoc for AI memories.

One canonical schema to bridge AI memory systems.

Stop writing n² converters between memory formats.
Write one Reader + one Writer per format. Everything else is automatic.

Explicit compatibility reports. No silent data loss.

Built for developers building AI assistants, memory platforms, and agent frameworks.

Why memlink?

Different AI tools store memories differently. memlink lets them exchange memories without every project writing custom converters.

Without memlink — 10 memory formats need 45 converters. Every new format makes it worse:

Ombre → OpenClaw    Ombre → Mem0    Ombre → Zep
Mem0 → Zep          ...

With memlink — each format writes one Reader + one Writer:

         ┌── Reader → Canonical ──┬── Writer → OpenClaw
  Ombre ─┤                        ├── Writer → Any Format
         └── ...                   └── Writer → Any Format

n systems = 2n plugins. Not n² converters.

Quick Start

pip install memlink-bridge

# Create a demo memory file
mkdir -p /tmp/memlink-demo
cat > /tmp/memlink-demo/hello.md << 'EOF'
---
title: Hello MemLink
tags: [demo]
---
This is a memory. memlink can bridge it across AI memory formats.
EOF

# Convert Generic Markdown → OpenClaw
memlink convert --from generic --to openclaw \
  -s /tmp/memlink-demo \
  -T /tmp/memlink-output

# Show installed formats
memlink formats

What just happened: A plain Markdown file was read into Canonical Memory format, then written as an OpenClaw memory.

Sample output:

Read:     1 memories from generic
Warnings: 0
Time:     0.00s

You should now see an OpenClaw-style memory file in /tmp/memlink-output/. The Compatibility Report tells you exactly what was preserved or dropped — no silent data loss.

Supported Formats

Generic Markdown works with YAML-frontmatter Markdown used by tools like Obsidian, Logseq, Bear, and plain Markdown. Tool-specific extensions are preserved as metadata or reported as compatibility notes.

Planned Formats

Feature Compatibility

MemLink is honest about what transfers and what doesn't.

Without a compatibility layer:

• Format-specific fields can disappear silently.

With MemLink:

• Preserved fields stay in Canonical Memory.
• Format-specific fields are preserved in metadata when possible.
• Compatibility reports explain what changed.

Multi-format demo — convert, merge, broadcast:

$ memlink convert --from ombre --to mem0 -s ./ombre-data -T ./mem0-out --verbose

Read:     4 memories from ombre

Compatibility Report:
  [ok] Preserved via metadata:
    Emotion fields (valence/arousal): 3 field values (75%)
  [~] Degraded:
    Unsupported memory kinds: 2 field values (50%)
Warnings: 2
Time:     0.00s

$ memlink merge --sources mem0:./mem0-out ombre:./ombre-data --to openclaw:./merged --verbose

Source mem0: 4 memories from ./mem0-out
Source ombre: 4 memories from ./ombre-data
Sources:  2 (mem0(4), ombre(4))
Total:    8 memories
Unique:   4
Resolved: 4 conflicts (strategy: newest)
Warnings: 0
Time:     0.00s

$ memlink formats

Format          Reader     Writer
-----------------------------------
chatgpt         yes        no
claude_export   yes        no
generic         yes        yes
mem0            yes        yes
ombre           yes        yes
openclaw        yes        yes
zep             yes        yes

Architecture

  Ombre ──┐
  Mem0  ──┼──→ Reader → Canonical Memory → Writer ──┬──→ OpenClaw
Generic ──┘                                          └──→ Ombre

Each format implements three methods:

class FormatPlugin:
    def read(path) → ReadResult       # Format → Canonical
    def write(memories, path) → []    # Canonical → Format
    def validate(path) → [Issue]      # Integrity checks

Add a new format = write one plugin. Zero changes to core code.

What memlink is NOT

• ❌ Sync engine — v1.0 is export/import only
• ❌ Memory database — Works with files, not APIs
• ❌ Embedding store — No vector search
• ❌ Knowledge graph — No traversal or inference
• ⚠️ Battle-tested — v1.0 API is stable, but not yet tested at scale (10K+ memories) or with production workloads. Use with that in mind.

Roadmap

Development

git clone https://github.com/velnori/memlink.git
cd memlink
pip install -e ".[dev]"

pytest tests/ -v          # 257 tests
ruff check python/memlink/       # Lint
mypy python/memlink/             # Type check

See CONTRIBUTING.md for how to add a new format.

License

MIT — see LICENSE.