Experience persistence across training runs — minimal overhead approach

[robotmem]

Hey,

I've been running long training sessions in Isaac Lab (Reach, Push tasks with
Franka) and one thing I keep running into: when I tweak the reward function or
change some domain randomization parameters and restart training, all the
"institutional knowledge" from the previous run is lost.

The replay buffer resets, obviously, but even high-level stuff like "these
reward weights converged faster" or "this DR range caused instability" — there's
no easy way to carry that forward.

I put together a small helper that logs episode-level summaries (not raw
observations — just aggregated stats) to a local SQLite db and retrieves
relevant past entries at training start:

from robotmem.isaac import RobotMemIsaacHelper

helper = RobotMemIsaacHelper(collection="isaac_reach")
# on_train_begin / on_episode_end / recall — all optional, non-blocking
# SQLite writes are batched, ~0.1ms overhead per episode

The main thing I wanted to make sure is that it doesn't slow down the training
loop. All DB writes are batched and happen outside the step() call. On a 4096-env
setup the overhead was <0.01% of total step time.

Has anyone else looked into cross-run experience persistence? Curious if there's
an existing approach in the Isaac Lab ecosystem I might have missed.

https://github.com/robotmem/robotmem