Open Dream-RSI

What is this?

Every AI agent pays the same tax over and over: it fails a task, works out what went wrong, and then throws that understanding away when the session ends. Open Dream-RSI keeps what was learned — without touching the model.

The LLM stays frozen. What improves is everything around it:

• the exploration policy (even re-written as Python code by the LLM itself, gated by replay before it is trusted),
• a library of verified solutions (recipes) reused as warm starts,
• and the world model the policy is scored against: a Discovery Tree of every attempt, enriched with replay-safe error-class facts by the Sentinel engine.

Between runs the system "dreams": it replays thousands of strategies over the recorded history in-process, at zero API cost, and keeps only what beats the incumbent. An open implementation of Recursive Self-Improvement through Evolving Worlds (Dream-RSI, Zheng et al., arXiv:2609.14858).

Zero runtime dependencies — Python 3.10+ stdlib only, ~11.2k lines of library code, ~5.4k lines of tests, no build step.

Not another memory plugin. Experience-memory systems record what the agent learned and reuse it as-is. ODR is an evolving worlds loop: every attempt is recorded into a Discovery Tree, and between runs the system dreams — replaying thousands of strategy variants against that recorded world. Nothing is promoted on the model's word alone: a candidate policy or recipe replaces the incumbent only after surviving counterfactual replay.

How it compares — stored state, how entries earn trust, runtime cost, against Mem0, Letta, Reflexion-style reflection, DSPy optimizers and skill libraries: docs/comparison.md.

📄 Preprint: Open Dream-RSI: An Open-Source Library for Recursive Self-Improvement Around a Frozen LLM, with Replay-Gated Learned Policies and a Curated Knowledge Base — P. Orwat, 2026. PDF · LaTeX source

Quick start

The odr console script ships with the package (0.3.9+); from a bare checkout use python3 -m open_dream_rsi or pip install -e ..

Point the dreamer at any OpenAI-compatible endpoint — or nowhere: the model is auto-detected from /v1/models, so a local vLLM/Ollama needs only the base URL.

export OPENAI_BASE_URL=http://127.0.0.1:8000/v1   # vLLM / Ollama / any gateway
export OPENAI_API_KEY=***                         # any token for local servers

No agent handy? The bare CLI runs the same loop: odr --memory .dream_rsi loop --tasks tasks.json --once (drop --once for a daemon). No key? python examples/live_loop_demo.py runs the whole cycle on an in-process mock server. (dreamrsi on PyPI is an unrelated project — this library is open-dream-rsi.) Per-host details, MCP blocks, troubleshooting: docs/integrations.md. The key-free walkthrough and the Docker image: docs/try-it.md.

How it works

                        ┌───────────────┐
                        │  Frozen LLM   │
                        └──────┬────────┘
                               │ proposals
              online loop      ▼      offline dream
        ┌──────────── DiscoveryTree ──────────── ReplaySimulator ──────────┐
        │                     │                          │                 │
        │              node.sentinel              counterfactual          │
        │              (Sentinel facts,           rollout scoring         │
        │               replay-safe)                    │                 │
        │                     │                         ▼                 │
        │                     └────→ Policy improvement ────→ next cycle  │
        │                                                                   │
   SENTINEL MODE (tool-execution layer)                                     │
        │          │             │                                          │
   error-class   nudge       permute-gate ──→ block tool calls             │
   ledger        (once/session)              past 80% of budget ───────────┘

1. Online execution — the agent attempts each task with the real LLM and a sandboxed verifier, appending every attempt to a per-task Discovery Tree (code, test feedback, score, one-line plan, and — when a Sentinel engine is wired — structured error-class facts and an explicit termination reason).
2. Offline dreaming — instead of paying for real-world rollouts, the agent "dreams" over the recorded history: thousands of strategy variants are scored in an in-process replay simulator at zero external-call cost. The best parameters, programs and recipes are persisted and steer the next cycle.

Beyond parameter-level dreaming, the loop closes section 3 of the paper ("dreaming with code"): each cycle the LLM may rewrite the exploration policy itself as a small Python program, choose_action(frontier, step). Candidates are statically validated (AST gate), executed only in a hardened subprocess sandbox, and scored by counterfactual replay rollout on the recorded tree. A candidate replaces the incumbent only on evidence — and the incumbent is re-scored on the current tree, not trusted at its stored score. A crashing or cheating policy can never break the loop: expansion falls back to the greedy baseline.

Plug in your favourite agent

The loop speaks MCP (Model Context Protocol). One config block and your everyday agent can queue tasks for the dreamer and pull back verified recipes — zero extra LLM setup: the dreamer resolves its own brain (env → local goose config → localhost vLLM), and asks the endpoint itself which model to use.

Tools are registered prefixed per client (e.g. Hermes mcp_open_dream_rsi_odr_run_once). No model calls tools on its own initiative — trigger them via a recipe's instructions:, a hook, or an explicit ask ("check odr_recipes before you try again"). The AGENTS.md prompting block that makes the loop proactive lives in docs/integrations.md.

Per-host setup — OpenCode, Goose (auto-setup script), Hermes (incl. Docker), Codex CLI, Claude Code, Cowork/claude.ai connectors, DeepSeek Harness — plus the zero-configuration brain chain and troubleshooting: docs/integrations.md.

Providers

Precedence is always: explicit overrides → environment → endpoint model discovery (first id from GET /v1/models) → preset defaults; a pinned model the endpoint no longer serves self-heals by re-discovering once. Cursor has no public OpenAI-compatible chat endpoint — the cursor preset targets a local proxy over the Cursor login, and since 0.3.0 the default zero-proxy brain is the cursor-agent CLI. Full prose and CLI install: docs/integrations.md.

Library-level use (no CLI):

from open_dream_rsi import (DiscoveryTree, DreamEngine, ReplaySimulator,
                            DreamAgent, LLMConfig, OpenAICompatibleClient)

client = OpenAICompatibleClient(LLMConfig.from_preset("local"))
tree = DiscoveryTree()
dreamer = DreamEngine(simulator=ReplaySimulator(tree))
dreamer.run_offline_optimization(iterations=100)
agent = DreamAgent(dreamer=dreamer, client=client)

What persists between runs (--memory dir)

Switches worth knowing: --no-policy-code, --no-knowledge, --no-thoughts, --no-judge ablate the four optional machinery layers; --evolution-stop (opt-in) freezes policy regeneration after an anytime-valid betting-martingale alarm — no promotion evidence left in a category stops the spending, the last promoted policy stays active; --budget N caps API calls per cycle (dreaming stays free).

Automated dreaming: when does the loop actually run?

The seam is dream_once() (CLI odr dream, MCP odr_dream, or a host hook). You do not set a schedule — the module decides the tier per call, so firing it after every task is safe and costs nothing extra when there is nothing new to learn:

A busy day dreams a few times, a quiet day once, an idle day never; concurrent calls collapse on a single-instance lock.

An idle fleet is never fully idle (wake self-test). By default (ODR_WAKE=1, or --wake) a FULL dream that finds tasks.json empty seeds one self-test probe — "write add(a, b)", verified by tests, one API call — so the configured endpoint is exercised even with no real work queued. odr metrics then shows live api_calls and a solve rate (category wake) instead of a zero-call heartbeat; opt out with ODR_WAKE=0 or --no-wake.

For a mass install the whole thing is two commands, no host glue:

pip install open-dream-rsi        # library + CLI + MCP server
hermes plugins install "https://github.com/patrykorwat/open-dream-rsi#odr-trigger"
hermes plugins enable odr-trigger                # restart to load the hook
# optional overrides:
hermes config set plugins.entries.odr-trigger.settings.provider local
hermes config set plugins.entries.odr-trigger.settings.base_url http://.../v1

Fleet deployments: many workspaces, one timer

A timer-based fleet is one runner called by one service. The shipped runner loops a glob of per-workspace memory dirs and fires one dream cycle per dir:

scripts/odr_fleet_dream.sh   # env-configured; see the header for ODR_* knobs

A systemd oneshot unit + timer (or the per-host service you already run) calls it; the dream module itself decides maintenance vs full per workspace:

# /etc/systemd/system/odr-dream.service  (Type=oneshot)
[Service]
Type=oneshot
Environment=ODR_PKG=/opt/odr/share/open-dream-rsi
Environment=ODR_WS_GLOB=/opt/odr/workspaces/*/odr-mem
Environment=ODR_PROVIDER=cursor-agent
ExecStart=/opt/odr/share/open-dream-rsi/scripts/odr_fleet_dream.sh

# odr-dream.timer — fires the unit on your cadence (here: every 2 h)
[Timer]
OnCalendar=*-*-* 0/2:17:00
[Install]
WantedBy=timers.target

What separates a learning fleet from a heartbeat-only one is carried by the runner's environment — two ingredients:

Since 0.3.11 the wake self-test closes the visibility gap: every FULL dream with an empty queue exercises the configured endpoint once, so a correct fleet reads wake: n/n and ~1 API call per workspace per ~6 h in odr metrics. A fleet stuck at api_calls: 0 across sweeps is a timer running bare dream calls — fix the table above, then diagnose one workspace with odr --memory <dir> metrics (the llm errors classes name the missing env).

Results (evaluation moved to docs/benchmarks.md)

Evaluation policy: performance claims live on exactly one benchmark — goose × TravelPlanner, public and scored by the benchmark's own evaluators; the scripted suites (bench, bench-policy) are deterministic machinery self-checks and carry no model-quality claim.

On the preregistered eval split (79 tasks, one model): plan delivery 18/79 (23%) cold → 74/79 (94%) with the call-cutoff gate (McNemar 56↔0, p=1.4e-17). Every undelivered cold episode died pinned at the call cap with the data already in hand — the binding failure is termination; the gate arm is fixed choice-removal, no learned content in those episodes.

The learned machinery is validated on the scripted suites; live replay also showed fresh prompt-injected lessons are hazardous by default (18/20 → 6/20 with imperative framing, declarative restatement back to 18/20) — the finding that led to removing prompt-lesson curation in 0.4.0. Full tables, Wilson intervals, protocol and self-check contracts: docs/benchmarks.md.

Learning vs lifecycle vs history

The repository keeps four planes strictly apart:

The ALM manages the four artifact types (policy_parameters, policy_program, recipe) through an explicit state machine (CANDIDATE → VALIDATED → ACTIVE → STALE/SUPERSEDED/QUARANTINED → ARCHIVED, terminal REJECTED), with hard rules: activation is only an explicit promotion call carrying the evidence; every state change appends exactly one immutable event before exposing the new state; the materialized view must always equal rebuild_artifact_state(events) (a hand-edited view is rebuilt from the log, never trusted); merges create NEW artifacts with lineage and never mutate sources; deletion never prunes the event log. Discovery Trees and audit events are evidence, not artifacts — the ALM refuses to register them. Inspect the store: odr artifacts [--type policy_program --state ACTIVE --history].

Security model

Candidate code — both task solutions and policy programs — runs in a hardened subprocess (open_dream_rsi/sandbox.py): python -I, scrubbed environment (only PATH — API keys never cross the wall), a timeout that kills the entire process group, and POSIX address-space / CPU / thread-count limits. Policy code additionally passes an AST gate before any spawn: required entry point, no imports, no dunder access, no __builtins__ reference (name plus constructed-key indexing is itself an escape route).

This is defence in depth, not a container. The subprocess does not restrict filesystem or network access, and an AST gate cannot be provably complete against code that is not statically gated. For hostile task sources set ODR_SANDBOX_CMD to a bubblewrap/nsjail wrapper (e.g. bwrap --unshare-all --die-with-parent --ro-bind / /) or run the loop inside a container.

On non-POSIX hosts (Windows) the same python -I child runs, but there are no rlimits and the timeout kills only the direct child — treat Windows as out of scope for hostile task sources. When ODR_SANDBOX_CMD is unset, each fresh process logs a one-time warning naming exactly this gap.

Live dashboard

One command, no dependencies, stdlib http server:

python -m open_dream_rsi dashboard                   # mock LLM, port 8765
python -m open_dream_rsi dashboard --provider cursor # real model
python -m open_dream_rsi dashboard --host 0.0.0.0    # LAN access

Dark single-page UI: KPIs (cycles / solved / API budget / dream iterations), live event feed, per-task attempt boards with score bars and code diff-downs, dreamed-policy gauges per category and the learned recipe library.

The dashboard is a plain unauthenticated stdlib HTTP server whose control routes (/api/pause, /api/reset) mutate the running loop — keep --host on loopback unless the LAN is trusted.

Positioning

• The paper this implements: Dream-RSI (arXiv:2609.14858); official repo zhengkid/Dream-RSI, explainer dream-rsi.com.
• The broader published line this sits in: Self-Refine (arXiv:2303.17651) iterates on outputs at test time; Reflexion stores verbal self-feedback across attempts; ExpeL (arXiv:2308.10144) distils cross-task insights and exemplars; Agent Workflow Memory (arXiv:2409.07429) induces reusable workflows; ACE (arXiv:2510.04618) evolves the agent's context as an accumulating playbook — the closest published analogue of our recipe store; Early Experience (arXiv:2510.08558) uses the agent's own rollouts as reward-free training signal (a weights-updating neighbour of dreaming); the field is mapped by the self-evolving-agents survey (arXiv:2507.21046, TMLR 01/2026). None of these promote candidates through counterfactual replay over an evolving recorded world — that promotion gate is this library's structural difference.
• Peer implementations (this repo is an independent open one, not the first): TheAstrayDev/dream-rsi-sdk (adapter SDK, strict replay; LLM-written policies on their roadmap), robinber/dream-rsi-spark (local Qwen + CUDA kernel exploration), plus skill variants (hermes-dream-rsi, dream-rsi-skill, Pi-RSI). What differentiates this repo: an always-on autonomous supervisor with budget guards, LLM-written exploration policies gated by counterfactual replay (section 3 closed), a public-benchmark measurement (goose × TravelPlanner) rather than self-scored claims, a deterministic decoy-trap self-check suite, persistent cross-run memory, and a negative-results section on memory that the loop itself enforces.
• OpenRSI / Frontis-MA1 (FrontisAI/OpenRSI) is a different layer: they post-train weights; this library optimises the exploration policy around a frozen model — no training, no GPU, no weight access. Complementary: their trained improver could be the frozen LLM behind our client.

Module architecture

Tests

python -m unittest discover -s tests      # full suite, ~2 min
python scripts/fast_tests.py              # fast subset (slow bench/sandbox modules excluded), ~15 s

Stdlib-only: benchmark behavioural contracts (greedy must solve nothing; the gated arm must dominate ε; the rollout must be prefix-only), sandbox escape regressions, sandbox result-cache identity contracts, holdout reservation semantics (first-claim-wins, cross-purpose refusal), gate semantics, MCP protocol, goose config dialects, sentinel channel semantics (nudge fires once per budget, merges with recurrence notes, clean calls count toward the budget; the permute-gate serves until ceil(0.8·budget) then refuses forever, per session; the 30-day production replay must reproduce its recorded note count exactly), model discovery and stale-pin self-heal contracts, and the structured seam (cross-session counts never reach to_world_dict(); the frontier-feature attribution pattern — beats greedy with metadata, ties without; the vendored Hermes engine pinned byte-for-byte to upstream by test_vendored_engine_matches_upstream).

License

MIT — see LICENSE.