Task traces, success/failure patterns, few-shot memory injection.

Episodic Memory — Learning from Past Agent Experiences

Episodic memory stores records of past agent experiences: task inputs, actions taken, results achieved, and whether the task succeeded. Agents retrieve similar past episodes to guide current task execution — a form of in-context learning from experience. Few-shot injection of successful past episodes improves performance on recurring task types. Error databases capture past mistakes and their resolutions, enabling agents to avoid repeating failures.

Interactive Pipeline Flow

UnderstandImplementOptimise

⚡ Execute — task run

Execute a task and capture the full trace: input, tool calls made, intermediate results, final output, and total tokens used.

Architecture Diagrams

Structure of an episodic memory record.

Loading diagram...

Key Design Decisions

💉Few-shot from past episodes beats system prompt alone

Injecting 2-3 successful past episodes as few-shot examples consistently outperforms asking the agent to "do its best." Real past traces carry more information than instructions.

🚫Store failures too — they are equally valuable

Failed episodes with their error messages teach agents what not to do. An error database (task→error→fix) injected into context can prevent the agent from repeating the same mistake.

⭐Score episodes with multiple signals

Combine automated checks (output format valid, task complete), user feedback, and LLM evaluation. Single-signal scoring misses important quality dimensions.

Discussion