prehend

A language-model harness that learns: recursive context offload, self-reflective program search, and experience memory.

prehend (to grasp - both comprehend and seize) is a harness that learns from the long-context problems it solves. It builds on rlms, the MIT OASYS lab's inference engine for Recursive Language Models (RLMs) - which replaces the canonical llm.completion(prompt) call with rlm.completion(prompt): the context is offloaded into a variable inside a REPL environment, and the model writes programs that slice, search, and recursively query that context instead of attending over it directly. prehend keeps that engine and adds the missing axis - a memory of what worked - so each solve makes the next one cheaper and better.

It layers three things on top of the upstream engine:

1. Experience memory. Completed solves are distilled into reusable bank entries, embedded, and retrieved on later tasks, so the harness carries forward strategies that worked instead of re-deriving them every run. This is the capability the prehend name is about (see docs/decisions/0005-prehend-experience-memory-layer.md).
2. Map-reduce style orchestration. Patches that harden the orchestrator-plus-workers pattern: long contexts are chunked and fanned out to parallel batched sub-calls (the map), and the orchestrator aggregates the partial answers (the reduce). Adds distinct system prompts for the orchestrator and its workers, per-child iteration budgets, and client fixes needed to drive local OpenAI-compatible servers reliably.
3. Self-reflective program search (SRLM). An SRLM subclass implementing uncertainty-guided trajectory selection per Apple's SRLM paper: generate K candidate context-interaction trajectories, then select using the model's own uncertainty signals (self-consistency, verbalized confidence, reasoning trace length) instead of trusting a single rollout. The same paper motivates context-length routing, since recursive decomposition often hurts when the context already fits the model's window.

Lineage

Stage	What it contributed
rlms 0.1.1 (Zhang, Kraska, Khattab)	The RLM paradigm and engine: REPL environments, recursive sub-calls, parallel rlm_query_batched, clients, logging, visualizer
Local rlms patches	Map-reduce orchestration support: child_system_prompt (workers get a different system prompt than the orchestrator), child_max_iterations, max_output_chars stdout truncation, default_extra_body on the OpenAI client, consecutive same-role message merging (required by llama-server), response_format pass-through
prehend	The SRLM subclass (context-length routing, multi-trajectory generation with parallel candidates, joint uncertainty-guided selection) plus the experience-memory layer that distills and retrieves past solves

SRLM: uncertainty-guided trajectory selection

The quality of an RLM answer depends heavily on which program trajectory the model happens to sample. SRLM subclasses RLM and replaces single-rollout inference with search over K candidates:

from prehend import SRLM

srlm = SRLM(
    backend="openai",
    backend_kwargs={"model_name": "my-model", "base_url": "http://localhost:8080/v1"},
    direct_threshold=30_000,      # contexts under 30K chars skip the REPL entirely
    n_candidates=4,               # K candidate trajectories
    candidate_parallel=2,         # candidates in flight at once (match server slots)
    candidate_temperature=0.7,    # sampling diversity across candidates
    confidence_elicitation=True,  # elicit per-step {"confidence": N} and use it in selection
)

result = srlm.completion(long_context, "What changed between Q3 and Q4?")

How a winner is chosen, per the SRLM paper:

1. Self-consistency. Final answers are clustered semantically (normalization plus word-boundary containment, so "42" and "The answer is 42" vote together) and the plurality cluster survives. Tied clusters pool their candidates rather than favoring whichever answer appeared first.
2. Joint uncertainty score. Within the surviving set, each trajectory gets VC(p) * Len(p), where VC is the sum of log per-step verbalized confidences (steps that skip reporting are imputed with the trajectory mean, so under-reporting cannot inflate the score) and Len is the trace length in output tokens. The candidate closest to zero wins. Without confidence_elicitation, selection falls back to the shortest trace.

Implementation notes:

• Each candidate runs on a fresh RLM instance with its own logger and config copy, so parallel candidates share no mutable state. A crashing candidate is dropped; only if every candidate fails does the call raise.
• confidence_elicitation=True appends the reporting instruction to the system prompt automatically; spawned candidates inherit it.
• direct_threshold routes short contexts to a plain LLM call. The SRLM paper finds recursive decomposition frequently underperforms the base model within its native window, so set this to roughly the served context size.

Parameter	Default	Meaning
direct_threshold	0 (off)	Context length in chars below which the REPL is bypassed
n_candidates	1	Candidate trajectories per completion
candidate_parallel	1	Candidates run concurrently (thread pool)
candidate_temperature	None	Temperature injected into candidate backends
confidence_elicitation	False	Elicit per-step confidence and use VC*Len selection

All RLM constructor arguments pass through unchanged, including child_system_prompt.

Install

Requires Python 3.11+. Available on PyPI; note that pip install rlms installs the upstream package, not this fork.

pip install prehend

For development, install editable from a checkout:

uv pip install -e /path/to/prehend --no-deps

Verify you got the fork and not a stale upstream build:

python -c "import inspect; from prehend import RLM, SRLM; print('child_system_prompt' in inspect.signature(RLM.__init__).parameters)"

Quick start

from prehend import RLM

rlm = RLM(
    backend="openai",
    backend_kwargs={"model_name": "gpt-5-nano"},
    verbose=True,
)

print(rlm.completion("Print me the first 100 powers of two, each on a newline.").response)

For the orchestrator/worker split used in map-reduce style runs:

rlm = RLM(
    backend="openai",
    backend_kwargs={...},
    custom_system_prompt=ORCHESTRATOR_PROMPT,   # the root model plans and reduces
    child_system_prompt=WORKER_PROMPT,          # sub-call workers map over chunks
    child_max_iterations=5,
    max_concurrent_subcalls=4,
)

REPL environments

Non-isolated environments run code on the host (fine for benchmarking, not for untrusted prompts); isolated environments run in cloud sandboxes. Natively supported: local (default), ipython, docker, modal, prime, daytona, e2b.

rlm = RLM(
    environment="local",
    environment_kwargs={"max_output_chars": 500},
)

• local: in-process exec with namespaced globals. max_output_chars truncates REPL stdout fed back to the model.
• ipython (pip install 'prehend[ipython]'): real IPython session, in-process or in an ipykernel subprocess with hard cell timeouts.
• docker: REPL inside a container (python:3.11-slim by default).
• modal / prime / daytona / e2b: fully isolated cloud sandboxes; sub-calls are proxied back to the host.

Model providers

OpenAI, Anthropic, OpenRouter, and Portkey clients are included. Local models work through any OpenAI-compatible server (vLLM, llama-server); the fork's default_extra_body and same-role message merging exist specifically to make local serving smooth. See prehend/clients/ to add providers.

Trajectory metadata and logging

RLMChatCompletion.metadata holds the full trajectory (run config plus every iteration and sub-call) when a logger is attached. SRLM relies on this for confidence scoring, and spawns per-candidate loggers automatically.

from prehend import RLM
from prehend.logger import RLMLogger

logger = RLMLogger(log_dir="./logs")   # omit log_dir for in-memory only
rlm = RLM(..., logger=logger)

JSONL logs feed the bundled visualizer:

cd visualizer/
npm run dev   # default localhost:3001

Citations

This fork builds directly on two papers. The engine:

@misc{zhang2026recursivelanguagemodels,
      title={Recursive Language Models},
      author={Alex L. Zhang and Tim Kraska and Omar Khattab},
      year={2026},
      eprint={2512.24601},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.24601},
}

The selection strategy:

@misc{alizadeh2026srlm,
      title={Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context},
      author={Keivan Alizadeh and Parshin Shojaee and Minsik Cho and Mehrdad Farajtabar},
      year={2026},
      eprint={2603.15653},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2603.15653},
}

Upstream documentation, blogpost, and minimal implementation: docs | blogpost | rlm-minimal.