Brainiak Topological Al drives LLM by loRa runtime adapters - Hackster.io<br>&times;Embed the widget on your own site<br>Add the following snippet to your HTML:

Brainiak core is a topological decision architecture that turns action results into next-step action capacity: statefull_infinite context<br>Read up about this project on

Brainiak Topological Al drives LLM by loRa runtime adapters

Brainiak

1 1

Overview<br>Things<br>Story<br>1. Brainiak first, Brainiak_Q second<br>Brainiak<br>Brainiak_Q<br>2. Non-negotiable architecture constraints<br>3. Canonical loop<br>4. Minimal proof: “weather in Lille” -> “the 1<br>Turn 1<br>Turn 2<br>5. Why this is not just RAG or classic tool-calling<br>RAG<br>Classic tool-calling agents<br>Chat-history bots<br>Brainiak<br>6. Runtime model: bounded active state, extensible memory horizon<br>7. CPU-first, on-prem, traceable<br>8. What is claimed (and what is not<br>Claimed<br>Not claimed<br>9. Screenshots to include (evidence pack<br>10. Theoretical foundation (arXiv preprints<br>Conclusion<br>Schematics<br>Code<br>Credits<br>Comments(1)

Share on FacebookShare on LinkedInShare on PinterestShare on RedditShare on XEmbed" data-html="true" data-placement="bottom" data-toggle="popover" data-trigger="click">

Brainiak

Published July 9, 2026 &copy; CC BY-NC

Brainiak Topological Al drives LLM by loRa runtime adapters<br>Brainiak core is a topological decision architecture that turns action results into next-step action capacity: statefull_infinite context<br>ExpertWork in progress125 days1

Things used in this project

Hardware components<br>AMD threadripper pro 7985 wx&times;1

Story

1. Brainiak first, Brainiak_Q second

Brainiak<br>Brainiak is the decision brain (ZTTM).It is designed for:<br>Decision sovereignty outside the LLM.<br>Stateful runtime with bounded active context.<br>Verifiable transitions through receipts.<br>Memory that becomes actionable, not just textual.

Brainiak_Q<br>Brainiak_Q is the governed LLM runtime layer on top of Brainiak.“Q” refers to Qwen 3.5 27B (+ LoRA adapters via vLLM).<br>Role split:<br>Brainiak/ZTTM decides.<br>ToolHub executes.<br>Brainiak_Q orchestrates governed runtime projection.<br>The LLM verbalizes and reformulates, but does not decide.

2. Non-negotiable architecture constraints<br>Brainiak is defined by strict invariants:<br>The LLM does not decide.<br>The frontend does not carry decision context.<br>Tools are not “intelligent.”<br>ToolHub executes but does not arbitrate.<br>ZTTM remains the decision authority.<br>Every action emits a receipt.<br>Receipts feed live memory.<br>Memory becomes operational (affordance-generating), not only conversational.

3. Canonical loop<br>Brainiak runs this loop:<br>perception -> ZTTM decision -> ToolHub action -> receipt -> live memory -> affordance -> next ZTTM decision -> next action<br>This is fundamentally different from:<br>conversation history -> LLM answer

4. Minimal proof: “weather in Lille” -> “the 1”<br>The scenario is intentionally simple.

Turn 1<br>User asks for weather in Lille.<br>ZTTM selects web_search.<br>ToolHub executes.<br>System emits a verifiable indexed receipt ([1], [2],...).<br>Receipt is written into live memory (memory_w, LoRA_context_layer, dynamic_knowledge_payload).

Turn 2<br>User says only: “the 1”.<br>ZTTM resolves this pointer from prior receipt state.<br>ZTTM selects http_request.<br>ToolHub executes and emits a second clean receipt (title, canonical URL, description).<br>Expected diagnostics in this flow:<br>llm_used=false on action decision in turn 2.<br>Route trace consistent with receipt-driven decision flow (for example: brainiak_q_zttm_toolhub_receipt).<br>Why this is a big deal:The UI interaction is tiny, but the architecture signal is huge:a prior event is transformed into next executable intent without LLM arbitration.

5. Why this is not just RAG or classic tool-calling

RAG<br>Retrieves context to improve text generation.

Classic tool-calling agents<br>Often let the LLM decide tool routing.

Chat-history bots<br>Mostly replay textual context.

Brainiak<br>Uses receipts as operational state:<br>Action result -> receipt.<br>Receipt -> affordance.<br>Affordance -> next ZTTM decision.<br>ToolHub executes under explicit decision authority.

6. Runtime model: bounded active state, extensible memory horizon<br>Brainiak does not claim infinite token context in one window.<br>It uses:<br>Bounded active runtime state (sliding discipline).<br>Persistent memory layers (Postgres/Qdrant via memory_w).<br>Controlled reprojection into live operational context.<br>So the right framing is:bounded runtime, extensible memory horizon.

7. CPU-first, on-prem, traceable<br>Brainiak is built around CPU-first conceptual/topological inference in the decision core.The LLM is integrated as a governed effect layer.<br>Practical consequences:<br>Better control over system behavior.<br>Explicit separation of decision/execution/verbalization.<br>Natural on-prem deployment posture.<br>Auditable action transitions through receipts.

8. What is claimed (and what is not)

Claimed<br>A validated receipt-mediated operational memory loop.<br>Decision sovereignty outside the LLM in the demonstrated flow.<br>End-to-end auditable transitions.

Not claimed<br>No AGI...