Cross-task experience retrieval for lifelong learning

[robotmem]

Hi,

I've been working with LIBERO-Spatial and one challenge I keep running
into: when training on a sequence of tasks, the model doesn't have
access to what it learned in earlier tasks. The lifelong learning setup
assumes knowledge transfer, but there's no built-in mechanism to
retrieve task-level insights across the sequence.

For example, after training on "open the cabinet", the experience that
"rotating the handle 15° clockwise first reduces jamming" could be
useful for "close the cabinet" — but that knowledge isn't persisted
anywhere queryable.

I wrote a small helper that logs task-level outcomes and retrieves
relevant past experience when starting a new task:

from robotmem import RobotMemory

mem = RobotMemory(collection="libero_spatial")
# after each task: mem.learn("handle rotation strategy worked", context=...)
# before new task: tips = mem.recall("cabinet manipulation tips")

It's just a local SQLite db — no changes to LIBERO or MuJoCo needed.
I tested it on simpler Gymnasium envs first (FetchPush-v4: ~42% to ~67%
over 300 episodes), but the cross-task retrieval seems most relevant for
LIBERO's lifelong learning scenario.

Has anyone else tried explicit experience persistence across the task
sequence? Curious whether this overlaps with something already in the
LIBERO pipeline.

https://github.com/robotmem/robotmem