REMEMBER.md Spec v1
Open standard for AI memory configuration. Status: Living Draft · Last updated 2026-05-05
Pitch
AGENTS.md tells your AI how to work. REMEMBER.md tells your AI how to remember. Together they form the markdown duo for portable, tool-agnostic AI configuration — one file for behaviour, one for memory, both human-readable and version-controllable.
Core rules
- The file is named
REMEMBER.md. - It lives at the root of your brain (or any project that needs overrides).
- It is plain markdown.
## Section Nameheadings delimit blocks. - Seven section names are reserved by the spec (see below). All optional.
- An empty file means "use engine defaults". Zero config = it works.
- Engines MUST ignore sections they don't understand (forwards-compatible).
- Engines MUST NOT auto-modify the file. It belongs to the user.
- User rules override engine defaults; project-level overrides global.
Reserved sections
| Section | Purpose |
|---|---|
## Capture Rules | Always-capture / never-capture filters and thresholds. |
## Routing | Where each kind of content lands in the brain. |
## Processing | How content is formatted, tagged, summarised. |
## Custom Types | User-defined entity types beyond the engine's defaults. |
## Connections | Auto-linking rules and relationship hints. |
## Language | Multilingual capture and processing preferences. |
## Promotion Thresholds | Confidence/sources thresholds for promoting beliefs to a pinned section. |
## Templates | Override engine-default templates per file kind. |
Override semantics
- Same-name section in your file — content is appended to the engine default.
## Override: <Name>— your content fully replaces the matching default.- Any other heading — passed through verbatim as a user-defined section. Engines ignore unknown sections gracefully.
Engine guarantees
- The brain is plain markdown, browseable in any editor (Obsidian, VS Code, …).
- Schema is additive:
type,freshness,confidence,sources_count,evidence[],counter_evidence[]are optional in v1, not required. - An evolution layer (consolidation, reflection, promotion) is allowed but must be auditable and reversible.
- No content leaves the user's machine without explicit consent.
Reference parser (20 lines)
import re
def parse_remember_md(path):
with open(path) as f:
text = f.read()
sections = {}
for match in re.finditer(
r'^## (.+)\n(.*?)(?=^## |\Z)',
text, re.M | re.S
):
name = match.group(1).strip()
content = match.group(2).strip()
if content:
sections[name] = content
return sectionsReference implementation
The Remember plugin (github.com/remember-md/remember) is the canonical engine. It ships:
- A SessionStart hook that loads
Persona.mdand an AI-driven passive capture nudge - A capture skill triggered by phrases like "remember this:" with live polarity check (positive evidence vs counter-evidence)
- A bulk-extraction skill for retroactive session processing
- Auto-promote at capture time so
Persona.md ## Top Beliefsstays in sync without cron, with bootstrap thresholds for cold-start brains - An
evolveskill (recommended weekly) that consolidates entities, reflects on beliefs, and promotes top beliefs to Persona - An auditable evolution log at
~/.local/state/remember/evolution.log
Adopting REMEMBER.md
Tool authors are encouraged to read REMEMBER.md from the project root and apply the rules described above. A 20-line parser (in any language) is enough for v1 compliance. Engines are free to extend with additional sections, as long as they do so under the ## x-* prefix or as ## Override: <name> blocks to avoid name conflicts.
Versioning
- The spec is versioned (v1.0, v1.1, v2.0); individual
REMEMBER.mdfiles are not. - New section names may be added in minor versions. Removed names: never.
- Engines stay forwards-compatible by ignoring unknown sections.
Status & contributions
v1 is a living draft. Implementations, feedback, and competing parser libraries are welcome at github.com/remember-md/remember/issues.
Inspired by .editorconfig (one page, dead-simple, 30+ editors), llms.txt (markdown-native, grassroots), and AGENTS.md (cross-tool agent instructions).