New Approach to Agent Memory

Skill Memory

Agent memory stored as skills — filesystem-compatible, configurable, and human-readable

Your agents build memory as plain markdown skill files you can read, edit, version control, and mount in sandboxes. No opaque embeddings. No vendor lock-in.

Three Advantages

Skill memory gives you properties no other agent memory system offers.

Filesystem-Compatible

Skills are real files on disk. Mount them in sandboxes, sync across environments, version control with Git, and inspect with cat and grep.

skills/daily-logs/
SKILL.md
2025-02-20.md
2025-02-21.md
2025-02-22.md

Configurable

Each SKILL.md defines the schema, purpose, and organization rules. You control how your agent stores memory — not the platform.

SKILL.md
# Daily Logs
description: One file per day
naming: YYYY-MM-DD.md
retention: 30 days

Human-Friendly

Plain markdown files your team can read, edit, review, and audit directly. No embeddings decoder needed — just open the file.

2025-02-22.md
## Deployed v2.1 to staging
Fixed auth timeout bug.
Updated rate limit to 1000/min.
User confirmed fix works.

How It Compares

See how skill memory stacks up against other approaches to AI agent memory.

Skill Memory (Acontext)
Vector StoreKnowledge GraphPlain-text Files
Storage formatMarkdown files
EmbeddingsNodes & edgesText files
Human-readable
Configurable schema
Complex upfront
Filesystem-native
Version controllable

How It Works

From completed tasks to reusable skill files — fully automatic.

1
Task Completes
An agent session finishes a task successfully.
2
Distillation
The learning pipeline extracts knowledge from the session.
3
Skill Agent
A specialized agent decides which skills to update and how.
4
Skills Updated
Skill files are created or updated as plain markdown.
SKILL.md
# Daily Logs
> description: Records daily activities and decisions
> naming: YYYY-MM-DD.md
> retention: 30 days

## Guidelines
- One entry per significant event
- Include outcomes and decisions
- Reference related tasks by ID
2025-02-22.md
## Deployed v2.1 to staging
- Fixed auth timeout bug (issue #342)
- Updated rate limit config to 1000 req/min
- User confirmed the fix resolves their issue

## Reviewed PR #89
- Approved with minor suggestions
- Performance improvement: 40% faster queries