chandra447/pi-hermes-memory

What it solves

Pi Hermes Memory is an extension for the Pi agent that provides persistent memory across sessions. It prevents the agent from forgetting user preferences, project-specific facts, and lessons learned from failures, ensuring that the agent doesn't repeat the same mistakes and remembers how to solve problems over time.

How it works

The project implements a hybrid memory architecture using Markdown files for human-readable storage and a SQLite database for fast, full-text search (FTS5). It manages three primary types of knowledge:

• Core Memory: Stores facts, preferences, and tool quirks in MEMORY.md and USER.md (capped at 5,000 characters each).
• Skills: Procedural knowledge (how-to guides) stored as SKILL.md files, which are unlimited in size and discoverable by the agent.
• Extended Store: A SQLite-backed mirror that allows the agent to search across all past conversations and extended memories using tools like session_search and memory_search.

To maintain this memory, the extension uses background learning (reviewing conversations every 10 turns or 15 tool calls), automatic correction detection (saving user corrections immediately), and a content scanner to block API keys and secrets from being persisted.

Who it’s for

Users of the Pi agent who want their AI assistant to maintain a long-term memory of their identity, their project conventions, and the processes they solve problems with, rather than starting every new session from scratch.

Highlights

• Two-Tier Memory: Separate global and project-scoped memory for facts that apply everywhere versus those specific to a codebase.
• Failure Memory: Specifically tracks what didn't work and why to prevent repeating errors.
• Procedural Skills: Allows the agent to save step-by-step procedures as reusable documents.
• Secret Scanning: Automatically blocks sensitive tokens and API keys from being saved to memory.
• Session Search: Full-text search across all past conversation history via SQLite.
• Standing Instructions: A dedicated, user-authored block of instructions that is injected into every session regardless of memory mode.
• Auto-Consolidation: Automatically merges and cleans up memory entries when character limits are reached.