memory-eternal

The biggest pain point of AI assistants is their goldfish memory: the moment a conversation ends, all context vanishes. Technical decisions, debugging insights, and project choices discussed last week are gone the next time you open a new window. Existing memory solutions either depend on third-party frameworks, lock memories into proprietary stores, or auto-ingest everything indiscriminately, letting noise and incorrect facts pollute future recalls.

memory-eternal takes a straightforward approach: at the end of each turn it distills useful content into concise knowledge cards persisted in a local SQLite database with zero third-party runtime dependencies. When an agent needs history it recalls only the relevant fragments, saving tokens and reducing noise. It serves DeepSeek Harness, Claude Code, Codex CLI, and Cursor simultaneously, all writing into one shared memory vault with per-agent attribution. A human-in-the-loop review center ensures only confirmed cards enter the main library, a 30-day soft-delete recycle bin provides safety, and a force-directed knowledge graph visualizes wikilink and shared-tag connections at a glance.

It suits three profiles: individual developers using DSH daily who want a talk-now-remember-later workflow; engineers juggling Claude Code, Codex, and Cursor who want a single shared memory layer; and budget-conscious users who can near-zero out LLM costs by disabling distillation and semantic dedup.