2SOLID AI — Synapse Brain
Every AI forgets. Ours doesn't have to.
Synapse Brain keeps your institution's context resident and queryable — memory the model interrogates, not fragments it hopes to retrieve. Powered by 4DAM, four-dimensional agentic memory.
Ask the brain something§01 — CLASSIFICATION
A memory substrate, not a memory feature.
Synapse Brain is infrastructure: one institutional memory that every agent in the stack reads and writes. The access pattern is interrogative — context stays resident and the model questions it — not retrieval of ranked fragments. Every returned record carries its evidence: source, verbatim quote, confidence, validity window.
§02 — ARCHITECTURE
Operating parameters.
| 01Designation | 4DAM | four-dimensional agentic memory |
|---|---|---|
| 02Access pattern | Interrogative | model asks; memory answers with citations — no ranked-chunk guessing |
| 03Interface operations | search · ask · ingest · capture | memory is a service the stack calls, not a log that accumulates |
| 04Evidence schema | text · sourceId · quote · confidence · validity window | carried per fact — every claim keeps its receipts |
| 05Scope | Tenant-isolated | scoped brains per organization — your institution answers your agents |
| 06Consumers | Synapse · AEGIS · BLUE PHYR · Cortex | one memory underneath the whole stack |
§03 — THE REJECTED ALTERNATIVE
Retrieval was the industry's answer. It was the wrong one.
RAG shreds documents into chunks and retrieves the top few — a filing cabinet the model glances at. Independent benchmarking keeps finding the same failure: retrieval bolt-ons can lose on accuracy to simply holding the context.
Retrieval (RAG)
Split it, embed it, hope the right chunks come back.
- Context arrives as fragments — ranked, truncated, guessed
- Each agent keeps a private notebook; nothing is institutional
- No provenance, no governance, no shared memory across a team
- When the right chunk isn't retrieved, the model fills the gap
Residency (4DAM)
Keep the context resident. Let the model ask.
- The brain holds the full context — the model interrogates it
- One institutional memory, governed, shared across agents
- Every claim carries evidence — source, quote, confidence, validity
- What one agent learns compounds into what the company knows
0
files opened to reconstruct a product's architecture
One ask_question call. File paths, symbols, and a commit hash returned as evidence.
That's what institutional memory buys: fluency without archaeology — for agents, and for every person who joins the company.
§04 — EVIDENCE
An agent interrogated the brain about a codebase it had never opened.
During the build of this site, an agent with no filesystem access to the Cortex repository issued one query. The brain returned its architecture as code citations — verified against commit f13359b and documented in the project record.
The Interview — ask_question("What is Cortex?")
Synapse Brain · ask_question · single call
- RESULT 1
Engine — mission queues, web fan-out (searchWeb / fetchPlain / brainSearch), spawned runs with depth caps, per-site concurrency limits.
cortex · src/main/research-engine.ts · Engine · commit f13359b
- RESULT 2
Fact — every claim carries text, a sourceId, a verbatim quote, a confidence score, and a validity window.
cortex · src/shared/research.ts · Fact
- RESULT 3
EditionSettings — brainControls + apiKeyCard: the Synapse Brain connection is baked into settings, not bolted on.
cortex · src/shared/editions.ts · EditionSettings
- RESULT 4
Shell — cortexapi bridge, assistant drawer with live execution trace, terminal, boot probes that ping the brain.
cortex · src/renderer/shell/Shell.tsx · Shell
- SYNTHESIS
Reconstructed: missions → runs → deliverables, an editions system, and brain-as-transport. Accurate enough that the person who built it recognized it — zero files opened.
§05 — FIELD INSTRUMENT
The banner above is a rotating 4-cube.
Sixteen vertices, thirty-two edges, double-rotating through four orthogonal planes, projected 4D → 3D → 2D while pulses ride the edges — the geometry the memory is named after, computed live.
| 01Object | Tesseract | the regular 4-polytope — 16 vertices, 32 edges, each vertex degree 4 |
|---|---|---|
| 02Edge rule | Hamming distance 1 | vertices adjacent iff they differ in exactly one coordinate |
| 03Rotation planes | XW · ZW · YZ · XY | 0.31 · 0.13 · 0.19 · 0.07 rad/s — incommensurate, so the figure never repeats |
| 04Projection | 4D → 3D → 2D | perspective at d₁ = 3.6, then d₂ = 7.0 |
| 05Edge classes | inner · outer · connector | w = −1 cube, w = +1 cube, and the 8 rungs between them |
| 06Traffic | 14pulses | data packets riding the edges, re-seeded on completion |
§06 — CAPABILITIES
What residency enables.
Shared, not per-agent
The brain is the institution's — governed, cross-agent, compounding. Not a notebook per hire.
Evidence-carrying
Memories keep their receipts — source, quote, confidence, validity. Ask a question, get back citations.
Queryable by design
search, ask, ingest, capture — memory is a service the whole stack calls, not a log that accumulates.
Onboarding, collapsed
What one agent learns on Monday, the workforce knows on Tuesday. Tribal knowledge stops walking out the door.
Tenant-isolated
Scoped brains per organization — your institution answers your agents, nobody else's.
Feeds the whole stack
Synapse personas, AEGIS operations, BLUE PHYR verification, Cortex missions — one memory underneath them all.
Ask the brain something.
Tell us what your institution needs to remember — we'll show you what resident memory answers.