I Built a Cross-Client Memory Hub for AI Agents — Here's What I Learned

Description and use

MemTether unifies 23+ AI clients into a single local SQLite database, eliminating cloud costs and improving context continuity. Benchmarks show 62.6% strict match and 54.6% LLM judge.

Problem

AI tools each maintain separate memory; switching tools causes loss of context.

Outcome

MemTether enables 23+ AI clients to share a single memory store, eliminating cloud costs and improving context continuity. Benchmarking shows 62.6% strict match and 54.6% LLM judge on LongMemEval, with multi-session improvements via E-Hybrid method.

Workflow

• Identify that all AI tools run on the same machine
• Create a local SQLite database to store memories
• Configure each AI client to point to the shared database
• Implement supersession logic to avoid deletion
• Add bi-temporal timestamps (valid and recorded times)
• Introduce Q-Value to rank frequently used memories
• Blend four re-ranking factors (semantic similarity, recency, usage frequency, type importance)
• Use SQLite triggers for FTS5 sync instead of Python code
• Package the engine as a pip installable wheel with console entry points
• Publish on GitHub and PyPI

Evidence

I built MemTether — a local-first memory hub that lets 23+ AI clients share one physical SQLite database.

Source

Existing memory solutions (mem0, cognee, zep) all treat memory as a service. You send memories to their cloud API, they store and retrieve them.

Source

No cloud. No API fees. No abstraction layer. One memory.db, 23 clients reading and writing to it.

Source

I ran LongMemEval (500 questions, full run): Strict match 62.6%, LLM judge 54.6%.

Source