Teach your AI agent to remember — and improve its memory over time.
SRSA (Spaced Repetition Systems for AI Agents) is a memory self-improvement layer that converts agent memories into reviewable cards and uses spaced repetition to continuously refine them.
Unlike traditional memory systems that only store and retrieve information, SRSA enables agents to:
- problem oriented enhancement of memory
- detect missing knowledge
- correct incorrect memories
- remove misleading memories
Most AI agent memory systems suffer from memory drift over time.
Common problems include:
-
Retrieval mismatch: Memories exist but cannot be activated in the right context.
-
Memory interference: Conflicting memories reduce recall accuracy.
-
Memory degradation: Memories become outdated, incomplete, or inaccurate.
These issues resemble findings in cognitive science such as the forgetting curve and retrieval failure.
Current approaches mainly focus on:
Pre-training approaches Designing better memory structures.
Post-processing approaches Summarizing or reorganizing memory.
However, they do not directly and effectively improve memory recall reliability during agent operation.
SRSA introduces spaced retrieval + memory self-correction to address this.
SRSA creates a continuous memory improvement process:
Card Generation → Scheduled Review → Self Evaluation → Memory Update
During review, the agent:
- recalls knowledge
- compares with ground truth
- evaluates correctness
- reflects on the mistake
- modifies memory
This enables continuous memory evolution.
Goal:
Next time: answer correctly and faster.
SRSA acts as a layer on top of any agent memory system.
Agent
↓
Memory System
↓
SRSA Skill
- Card Generator
- FSRS Scheduler
- Review Loop
- Memory Self-Correction
↓
Improved Memory
Key design principle:
-
Memory is dynamic: Agent knowledge should evolve.
-
Recall is the real bottleneck: Storage alone does not solve memory problems.
-
Review enables learning: Periodic retrieval strengthens memory pathways.
-
🎬Scenario-based recall improvement: Custom cards help solve retrieval failures.
-
🔓Decoupled from memory systems: Works with any agent memory architecture.
-
🚀Efficient scheduling: Uses the FSRS algorithm to minimize unnecessary reviews and token usage.
-
🔧Memory self-repair: Agents update memory after failed recall.
-
📈Review analytics: Tracks recall performance and memory evolution over time.
It is easy to integrate SRSA into your agent workflow using skills.
Install the skill srsa:
clawhub install srsaAsk the agent to read the skill, generate a card from its memory, and set a daily refresher task.
scripts/
main.py
config.yaml
cache/
fsrs.sqlite
doc/
script.md
SKILL.md
README.md
Optional configuration: scripts/config.yaml
You can refer to FSRS documentation for details.
Refer to script.md for command usage.
If daily reviews become large, recommended solutions are:
- lower
desired_retentioninconfig.yamlto reduce review frequency - compress conversation context
- split reviews across multiple agents
- run multiple review sessions per day
Q: What types of news does the user like?
Agent Answer:
Tecknology and science news.
Correct Answer:
Tecknology, science and international news.
Reflection:
Related memories are incomplete and need to be updated to include international news.
Action:
Update memory with the new information about international news.
MIT License