Does anyone also face repeated AI research across tools?

Posted by Vedant_d_@reddit | LocalLLaMA | View on Reddit | 2 comments

I work with multiple AI tools on same project, and I keep seeing this issue.

Tool A already explored context, but Tool B starts same research from zero again.

So I was losing time and tokens for repeated work.

I made a local setup to reduce this problem.

Simple flow:

one assistant saves findings
another assistant reads and continues
memory carries across sessions, so if work was done in previous session, next model can recall what was completed and continue from there
it can also recall useful project details from earlier sessions instead of re-discovering everything

It is local retrieval (FAISS + sentence-transformers), and I also got it working with CLI agents now (Qwen CLI, Codex CLI, Gemini CLI), not only IDE tools.

I am also testing multi-agent collaboration. It works, but I am still optimizing speed and handoff quality.

[-]

Icy_Host_1975@reddit

the pattern you described is exactly what a runtime with persistent memory does natively. OpenClaw (openclaw.vibebrowser.app) does this out of the box - memory carries across sessions and any new agent picks up where the last left off, no custom FAISS layer needed. the handoff quality issue is usually the hardest part - curious how you are structuring session boundaries between your CLI agents.h

Vedant_d_@reddit (OP)

Yes, I checked openclaw it offers per-tenant persistent runtimes and MCP endpoints. the project i'm working on is bit different, its a local memory layer that links IDEs and CLI agents (qwen/codex/gemini/claude, cursor, vscode) so local tools can share and reuse findings

for session boundaries, i track agent identity (model name + unique ID), so each CLI tool is isolated. Each agent explicitly joins a session, and verifies membership before reads/writes. the memory injection works across sessions

And about handoff quality, that's what I'm optimizing now. Right now agents need to manually call session_start/end; ideally it's more automatic on first MCP connect