Show HN: Lossless-memory – a personal AI memory that never summarizes
summary
Most AI memory systems fail by summarizing conversations, which strips away the essential context and nuance of human interaction. This project proposes a 'lossless' approach that stores every single line of conversation verbatim, using a temporal backbone to prioritize chronological retrieval over semantic similarity. By treating time as the primary search axis and avoiding summarization entirely, the system preserves the authentic history of a user's relationship with their AI.
key points
- The system utilizes a Temporal Backbone where timestamps are the primary index, allowing users to query by specific time ranges like 'last Tuesday night' to retrieve exact, unsummarized logs.
- A specialized 'where are we now' index called LLL provides the model with constant context of the current topic, ensuring the AI maintains its place in a conversation even after context window compaction.
- The architecture prioritizes raw JSONL logs as the absolute source of truth, using SQLite FTS5 for exact search and semantic vector search only as a last resort.
who should read it
Developers and AI enthusiasts interested in building personal, local-first long-term memory layers for LLMs that prioritize data integrity over compression.
discussion positive
consensus
Commenters generally view the approach as a useful information retrieval method for agentic memory, though they debate the specific utility of timestamps. While some argue that temporal ordering is essential for rebuilding operational state and handling evolving rules, others believe existing model capabilities or simpler markdown lists suffice. There is a shared skepticism regarding the long-term scalability of single-memory techniques, suggesting that true memory requires a complex, multi-layered system.
top insight
The most valuable perspective is that 'memory' in AI is not a single problem but ten different problems requiring distinct solutions. This insight suggests that any single approach, regardless of how clever, will eventually hit a wall and necessitate the addition of further memory techniques to meet diverse user needs. This shifts the focus from finding a 'perfect' memory system to building a modular, multi-layered architecture.
Generated by an LLM from the article and thread. Check the sources before relying on it.