Brain Agent

Neural-inspired knowledge retrieval system built on heaven-base.

Overview

Brain Agent provides a sophisticated system for organizing and querying knowledge from document collections. It uses a "neural" metaphor where documents become "neurons" that can be activated based on relevance to queries.

Components

• BrainAgent: Main agent class for knowledge retrieval
• CognizeTool: Identifies relevant neurons for a query
• InstructTool: Generates instructions from activated neurons
• QueryBrainTool: Simple interface for querying registered brains
• SynthesizerReplicant: Alternative replicant-based interface

Installation

# Install heaven-base first
pip install git+https://github.com/sancovp/heaven-base.git@v1.2.0

# Install brain-agent
pip install -e .

Usage

from brain_agent import BrainAgent, register_brain

# Register a brain (document collection)
register_brain(
    directory="/path/to/documents",
    brain_name="my_knowledge_base",
    chunk_size=-1  # whole files
)

# Create and query brain agent
agent = BrainAgent()
result = await agent.query("brain=my_knowledge_base query=What is machine learning?")

Dependencies

• heaven-base>=1.2.0 (for core agent framework)
• langchain-core (for message types)
• Various LLM providers (OpenAI, Google, etc.)

License

Private - All rights reserved.

Hierarchical Brains (v0.2.0)

Brains whose neurons are brains. A Brain implements the same 2-stage Neuron protocol its own neurons use — cognize (cheap relevance vote off the brain's _digest.md) and instruct (a full recursive cognize→instruct→synthesize pass) — so hierarchies nest to arbitrary depth. Digests build bottom-up (build_digests): the level touching raw files gets an LLM fold; levels above concatenate child digests verbatim, so distinctive vocabulary survives to the root.

from brain_agent.hierarchical import Brain, build_digests
await build_digests(Path("corpus"))            # the pyramid is just directories
answer = await Brain(Path("corpus")).query("...")

HTTP server (judge + fill)

brain-agent-http serves the brain over HTTP for coordinate/configuration engines: the caller owns addressing (which parts exist, which slots are empty); the server owns judgment and generation.

• POST /judge — {parts, rule} → one verdict per part (complies / violates / not_applicable) with a verbatim witness quote, machine-verified to appear in the source (witness_verified). Exhaustive: every part judged, none skipped. Returns a global_section flag (no violations anywhere).
• POST /fill — {slot_label, siblings, n, brain_root?} → candidate spectrum completions; if brain_root is given the proposals are grounded in a witnessed brain synthesis (grounded flag reports honestly).
• POST /brains/build, POST /brains/query, POST /neuron/cognize, POST /neuron/instruct, GET /health.

Runs entirely on heaven (UnifiedChat + HeavenAgentConfig) — heaven owns model routing and auth, exactly as brain_agent/tools.py does. Model is HBRAIN_MODEL (default MiniMax-M2.7-highspeed, provider HBRAIN_PROVIDER, default ANTHROPIC — heaven auto-routes MiniMax-* to MINIMAX_API_KEY). No raw provider SDK.

RLM — Recursive Language Model (v0.3.0)

RLM owns a GROWING corpus end-to-end: live ingestion, incremental pyramid maintenance, and query/judge over unbounded context. The context window is replaced by a filesystem pyramid — each LLM call sees O(node) tokens while the corpus is unbounded.

from brain_agent.rlm import RLM

r = RLM("corpus_root", session="my_session")
r.ingest_message("user", "...")        # boundary rule: user text starts an iteration
r.ingest_message("assistant", "...")
await r.reindex()                        # folds ONLY dirty branches — O(changed)
res = await r.query("...")              # witnessed synthesis + descent refs
report = await r.judge("<rule>")        # exhaustive incidence row over all parts

HTTP: /rlm/ingest, /rlm/query, /rlm/judge, /rlm/flush (stateful sessions keyed by root+session on the same brain-agent-http server).

Verified on a real 329-message agent transcript: 22 iterations / 3 phases auto-folded, incremental growth re-folds exactly the dirty branch, needle query answered with a verbatim source-cited quote.

The orchestrator — RLM as the paper defines it (v0.8.0)

The hierarchical brain above descends by a fixed Python program: the model only emits a relevance score and hierarchical.py decides what to open. The RLM paper (arXiv:2512.24601 §2) requires the inverse — a symbolic handle to the data, symbolic recursion (the model invokes sub-calls from code it writes), and the answer returned from a variable rather than a finish action.

A brain is N neurons + 1 synthesizer. In an RLM the synthesizer stops being a one-shot fold and becomes the thing that runs the turn: it builds brain systems and calls them however it wants, as many at once as it wants, until it is satisfied.

from brain_agent import Orchestrator

orch = Orchestrator(kernel="work")
orch.bind(contracts=big_string)          # anything, by name, into the context
answer = await orch.run("review these from three independent perspectives")

• One persistent runtime context. Neurons, brains and whole systems it builds are live Python objects in a kernel; results pipe between them as values, never re-serialized through the model.
• The shell is a real tool (ShellTool), called through heaven's native tool loop. The model's output formatting is not load-bearing.
• No turn cap. It decides when the answer is good enough. The only bound is a runaway backstop (BRAIN_MAX_TOOL_CALLS, default 1000).
• Recursion needs no special case, because a Brain is usable as a neuron.

A real run, from the call graph: the orchestrator built three lensed brains, ran them concurrently inside one shell call, then built a fourth brain whose neurons were those three analyses:

orchestrator Review these contracts from three independent expert  166.89s
  shell    import asyncio                                           43.91s
    brain    Legal Risk Brain            opened=3/3  scores=[10, 9, 9]
    brain    Financial Exposure Brain    opened=3/3  scores=[10, 9, 9]
    brain    Renewal Timeline Brain      opened=3/3  scores=[8, 9, 6]
  shell    from brain_agent import Neuron, Synthesizer, Compose     61.68s
    brain    Priority Synthesizer        opened=3/3  scores=[10, 9, 10]

Call-graph tracing (v0.6.0)

Every agent call is recorded as a node — root turns, cells, brains, votes, neurons, fanouts, sub-LLM calls, synthesizers, and sub_rlm delegations. Recording is off unless BRAIN_TRACE_DIR is set, and needs only the standard library; networkx is required only to load a graph (pip install brain-agent[graph]).

BRAIN_TRACE_DIR=/tmp/run1 python my_rlm_script.py
python -m brain_agent.trace /tmp/run1

root     Exactly one ledger entry is disputed. Which account,  10.51s
  turn     turn 0                                                 4.39s
  turn     turn 2                                                 2.69s
  cell     import re                                              0.00s
  cell     Final = {                                              0.00s

from brain_agent import load_run
import networkx as nx
G = load_run("/tmp/run1")
nx.is_tree(G.to_undirected())        # True — one caller per call
G.nodes["tr:4"]["elapsed_s"], G.nodes["tr:4"]["kind"]

The shape is a tree, but it is a DiGraph because it spans processes: a sub_rlm child runs in its own kernel and writes its own file. The parent's sub_rlm node carries the child's kernel name, and load_run() stitches on that edge (marked cross_process=True). Nesting inside a process is automatic via a contextvar, and the cell request carries the caller's span id so kernel-side work hangs under the client's tree rather than orphaning.

Nodes carry kind, label, elapsed_s, ok/error, plus extras: a vote node records the scores it produced, a brain node how many neurons opened.

Parallelism

• fanout(chunks, q) and Brain.vote issue N model calls in one batch — measured 2.6× faster than the same six calls sequentially.
• Brain.query runs opened files and sub-brains concurrently.
• sub_rlm awaits its child, but asyncio.gather (exposed in the shell as gather) runs several children at once — each on its own kernel.

Lenses (v0.7.0)

A neuron's prompt= is its lens — what it reports and what it ignores. Give the same content different lenses to get readings that cannot contaminate each other, then fold them with a Synthesizer whose prompt controls the form:

lenses = {"legal":    "Report ONLY legal risk. Ignore money and dates.",
          "money":    "Report ONLY amounts and payment terms.",
          "timeline": "Report ONLY dates and durations."}
ns = [Neuron(content=doc, name=k, prompt=v) for k, v in lenses.items()]
reads = await fanout(ns, "review this contract")     # each keeps its own lens
report = await Synthesizer(prompt="Return JSON with keys legal, money, timeline.")(
    dict(zip(lenses, reads)))

cognize_prompt= is the separate routing lens (how a neuron scores its own relevance). Lens separation quality tracks lens prompt quality: one-line lenses ground correctly but partition weakly; a lens that names what to ignore partitions cleanly.

CONTEXT — the state dict (v0.10.0)

The synthesizer agent's working memory is CONTEXT, a state dict in its shell. Each slot is a chunk of context, persisted to a file the moment it is set — and neurons read files, so a slot is directly callable context. A directory of slot files is a brain, so CONTEXT.brain() turns the working context itself into one, and register_brain(CONTEXT.dir, name) makes it permanent.

CONTEXT["contract_a"] = chunk               # persisted to <ctx>/contract_a.md
n = Neuron(content=CONTEXT.path("contract_a"), prompt="Report only risk.")
CONTEXT["risk_a"] = await n("review")       # results become context too
FINAL = report                              # ends the work; returned unbounded

The instructed loop: get material into slots → make vars and pipe them through neurons/brains reading the slot files → write results back as new slots → reslice and relens until satisfied → set FINAL. Slots survive kernel restarts (chunk files re-adopt on boot), bind() places values into slots, and FINAL is retrieved from the kernel so the answer is not bounded by the model's reply.