AI Memory Booster

Problem

memory-booster addresses two common memory failures in long-running AI agent work: new sessions losing prior progress, and long conversations losing earlier decisions, file paths, and follow-ups as tool calls accumulate. It turns “remembering” into a structured retrieval, compression, and archive layer around MEMORY.md, journal files, and historical chat context.

How It Works

The workflow centers on four operations:
- Warm-up: at session start, it extracts active topics from MEMORY.md, recent journals, and keywords, forms natural-language queries, and combines keyword search with semantic vector recall.
- Compression: it extracts important information into markers such as DECISION, DATA, LINK, ISSUE, NEXT, and FILE, appends to journal files, and updates only relevant sections of MEMORY.md.
- On-demand recall: when the user asks what was said before or provides a custom query, it performs semantic search first, supplements with conversation search, and returns only the most relevant entries.
- Industry skill generation: !forge generates a customized memory skill with industry-specific categories, extraction rules, keyword mappings, ignore rules, and compression templates, producing a standalone, distributable skill.

It is intended for WorkBuddy-style environments with file tools, conversation_search, and local Python/Bash scripts. Semantic indexing must be built once, memory archiving is manually triggered, and some in-chat context overflow still depends on platform-side support.