Memory and Exploration#

Memory lets runs reuse task experience; exploration builds local experience through repeated attempts. Complete Quick Start before following these LIBERO examples. See each environment’s page for its task arguments and success criterion.

Evaluation reads memory without updating it. Exploration can reset and retry. Formal reproduction must use the specified memory revision and evaluation mode.

Use Published Memory#

Evaluation defaults to --memory-profile hf and downloads memory from the public RLinf/RPent-memory dataset. LIBERO selects one version with --memory-version auto (the default):

LIBERO memory versions#
Running model	Memory directory under `libero/`	Exploration configuration
`gpt-5.5`	`GPT_5.5_xhigh`	Codex, reasoning on, xhigh
`gpt-6-astra`	`GPT_6_astra_low`	Codex, reasoning on, low

Provider prefixes such as openai: are recognized. Codex uses --model first, then CODEX_MODEL. Unknown models, Claude, or an unknown backend default fall back to GPT_5.5_xhigh with a warning. Flash replay defaults to GPT-5.5. An explicit version overrides model selection; it does not change the running model or reasoning effort. The effort in a directory name records how that memory was generated.

# Choose Astra memory automatically.
rpent --robot libero --suite libero_goal_swap --task 1 --seed 1 \
  --planner codex --model gpt-6-astra --reasoning-effort low

# Use the same model with the GPT-5.5 corpus.
rpent --robot libero --suite libero_goal_swap --task 1 --seed 1 \
  --planner codex --model gpt-6-astra --reasoning-effort low \
  --memory-version GPT_5.5_xhigh

CLI and Dashboard resolve the root before each task. In Dashboard, Next task model changes the model for the next task and reselects auto memory then; the active task keeps its existing model and corpus. A manually selected memory version remains selected across model changes.

Only the chosen version is downloaded. LIBERO caches are isolated by repository, commit and version under memory/libero/.versions/. Both the exact file set and every file hash are verified before cache reuse. Extra files invalidate the cache, including under a pinned revision. Online sync rebuilds an invalid cache; HF_HUB_OFFLINE=1 requires a complete, unchanged cache for the selected version and revision; failed downloads never substitute another model’s corpus. Missing or incomplete caches fail explicitly. Caches created before versioned-source receipts require one successful online refresh; old unversioned caches are not reused. Other robots retain their existing synchronization behavior.

Standalone Download and Local Evaluation#

python -m robots.libero.memory sync --memory-version GPT_6_astra_low
python -m robots.libero.memory sync --model gpt-6-astra \
  --revision <release-commit> --output-dir /path/to/new-astra-memory
rpent --robot libero --suite libero_goal_swap --task 1 --seed 1 \
  --planner codex --model gpt-6-astra --reasoning-effort low \
  --memory-profile local --memory-dir /path/to/new-astra-memory

sync prints the actual corpus root. --output-dir must not already exist. --planner defaults to api, matching rpent; pass --planner codex to use CODEX_MODEL when --model is omitted. --memory-profile local never downloads memory; combining it or --explore with an explicit remote --memory-version is an error. Exploration uses a local corpus; use a separate empty --memory-dir for each independent exploration.

Versioned downloads are a LIBERO-specific command. The shared rpent-memory command continues to provide merge, validate and build-index.

Release Provenance and Compatibility#

The dataset’s libero/README.md and libero/manifest.json document the versions, original source snapshots and published files. Each version’s files mapping contains paths relative to that version’s root and SHA-256 hashes of the published bytes; the loader verifies this mapping. Source revisions and source hashes describe the original snapshots and remain separate from published hashes after directory, index and reference changes.

Both version roots use MEMORY.md, global/, task-family/ and task-specific/. GPT-5.5 additionally includes 78 plan/anchor pairs in flash/, copied unchanged from the merged Flash release. Flash reads that directory within the selected corpus. Astra has no replay assets; selecting it for Flash reports an error.

The Astra release merges Long and Spatial/Object/Goal exploration memory, preserving both versions of three conflicting global notes with source suffixes. Its 79 task-specific audit/recipe pairs retain their original content; Long Swap task 6 has no task-specific pair. The historical 741/800 result used the two original frozen snapshots separately by suite. The merged release has not been reevaluated. Memory was generated at runtime commit 014a0fa, before the scene-seed fix. The original source revisions are cf5d14ce9b3ec6c72de5477ec0fea806a3884efa (Long) and 984c57f7c6caf48b572f5926851dd9134ae041d8 (Spatial/Object/Goal).

The current loader requires a versioned Hub layout and does not convert or fall back to the historical unversioned corpus. Update code and data together. The current directory names require the shared memory naming update in RPent #190 and the corresponding dataset update. Dataset main evolves; reproduce/memory keeps its historical contents and layout for the matching historical robot reproduction branches. Historical GPT-5.5 reproduction uses the matching historical client and the original dataset revision 21a62795fe3b7e500c8381ac47938f6d713ebe18:

hf download RLinf/RPent-memory --repo-type dataset \
  --revision 21a62795fe3b7e500c8381ac47938f6d713ebe18 \
  --include 'libero/*' --local-dir /path/to/legacy-download
# With an older RPent client:
rpent --robot libero --suite libero_goal_swap --task 1 --seed 1 \
  --planner codex --model gpt-5.5 --memory-profile local \
  --memory-dir /path/to/legacy-download/libero

Explore and Build Local Memory#

This command uses a separate local directory, with up to 3 planner sessions and 5 attempts per session. The directory may start empty; --explore selects the local profile. Choose a new output directory for later runs:

rpent --robot libero --suite libero_10_task --task 0 --seed 0 \
  --planner claude_code --model claude-opus-4-8 \
  --explore --explore-sessions 3 --explore-attempts-per-session 5 \
  --memory-dir ./memory/libero-local \
  --output-dir logs/explore_10_task_t0_s0

Each session creates a fresh toolkit; attempts within that session start with an environment reset. State and observations are saved under <output-dir>/sessions/session_NNN/, and experience drafts under <memory-dir>/_internal/inbox/<cell>/.

On normal completion, the runner validates and merges drafts and updates the index. It publishes successful task audits and action sequences only when LIBERO reports success. Generated data stays local and is not uploaded automatically. Add --dashboard to watch exploration; see Interactive Usage for controls.

Evaluate with Local Memory#

After exploration, inspect the local corpus and run result, then evaluate another scene with that corpus:

rpent --robot libero --suite libero_10_task --task 0 --seed 1 \
  --planner claude_code --model claude-opus-4-8 \
  --memory-profile local --memory-dir ./memory/libero-local

The local profile neither downloads memory nor enables exploration. LIBERO checks that a corpus exists and rejects an empty directory. The hf profile cannot be combined with --memory-dir.

Inspect memory-read tool calls in the transcript to confirm they use files permitted for the current task. Historical coordinates, pixels, and poses are references; actions still require localization against current observations.

Validate and Merge Manually#

Check memory file structure and rebuild the index from existing experience:

rpent-memory --memory-dir ./memory/libero-local validate
rpent-memory --memory-dir ./memory/libero-local build-index

To review drafts before publishing, add --no-auto-merge-memory to exploration. After review and verification of LIBERO’s success result, merge that cell with the command below. --solved is a manually supplied success flag; the planner’s summary alone does not justify it:

rpent-memory --memory-dir ./memory/libero-local merge \
  --cell 10_task_t0_s0 --output-dir logs/explore_10_task_t0_s0 --solved

Structural validation does not prove real task success. See Memory Design for layout, permissions, and publication rules.

Other Environments#

RoboCasa365 and RoboTwin support exploration; see their pages for task arguments, resets, and memory scope.
Dual-Arm Franka exploration requires an operator to reset the scene and judge results. Automatic merging is disabled by default. Complete deployment and motion checks first.
Single-Arm Franka does not currently support exploration.