A memory & organization substrate  /  research architecture

Identity and memory are topology,
not archive.

TEPiD stores where things are and how they relate, not the things themselves. Wrap a small agent in it and that agent reaches a long-horizon coherence, and a corpus access, you would normally need a far larger one to get.

scroll
The thesis, in one line

You don’t store what something is. You store that it exists, and roughly where to find it.

The detail is reconstructed on demand. A point’s meaning is its position in a relational cloud, and the gaps between points carry as much information as the points themselves. The system behaves like an index rather than a transcript: it knows where to find things without holding the things.

This inverts the usual route to a smarter agent. Most work makes the model bigger or the context longer. TEPiD keeps the agent small and its context tiny, and pushes the heavy lifting down into the substrate: memory, structure, decay, classification, retrieval.

Temperature is the spine of the design, read from creation rather than dissolution. A pointer is born hot, undifferentiated and maximally active, then cools into structure. The cold regions are the settled, individuated terrain, so here cooling does the work decay usually undoes. As a pointer cools, it takes shape. The field above is that process: sparks ignite hot, drift, and precipitate into a lattice as they go cold.

TEP: Topological Experiential Persistence.
iD: the durable, identity-bearing schema.
tepid: the state the system is always in. Everything has a temperature.


The load-bearing claim
“Small, cheap, local agents wrapped in TEPiD operate with long-horizon coherence and corpus access that otherwise requires much larger agents.”

The substrate does the work the agent shouldn’t.

The agent’s working context only ever holds the small slice currently on the table. Everything else stays resident in the substrate, which behaves like a library. It remembers everything it has been told, reorganizes itself as the agent learns, gets faster the more it is used, and forgets the unimportant gracefully while keeping the audit trail.

Neither side runs the other. The substrate asks the agent for judgment: classify, judge, summarize. The agent asks the substrate for memory: read, walk, write. They meet at an API and nowhere else.

  • For a frontier agentTEPiD is a coherence multiplier. Sharper judgment over a memory that never silts up.
  • For a small local agentTEPiD is a capability multiplier. A 2B-parameter model becomes useful well beyond its unwrapped reach.
  • It is not a modelTEPiD doesn’t reason; an agent does.
  • It is not vector RAGretrieval is geometric and addressable, not embedding-similarity search.
  • It is not a databaseit has a specific runtime shape: catalog, cache, content, harness, agent.

The substrate · three resident layers and a background runtime

The addressing layer stays resident. The content stays in the stacks.

The catalog holds tiny pointers into the content. The stacks hold the content itself. The embers hold what was just resolved, still warm. The walker consults the catalog, pulls what it needs, reads it, and releases it, while the harness reorganizes continuously in the background.

◇  THE WALKER  ·  any agent, any model, any size
↓   API: read · walk · classify · judge · summarize
L1 · catalog

The catalog

Resident, fixed-width sparks at 128 bytes each. The addressing layer is always hot.

1 B sparks ≈ 128 GB resident
L1.5 · embers

The embers

Recently-resolved content, still glowing. Microsecond access, transparent to the agent.

cache · (spark_id, generation, blob_hash)
L2 · the stacks

The stacks

Petabyte-scale content archive. Hash-addressed, multi-resolution, tiered hot to deep.

full / summary / stub · hot · warm · cold · deep
↑   the harness ticks here, decaying heat, verifying checksums, maintaining indexes, never blocking the read path

The geometry

Nested cubes. Relations read from position, not stored as edges.

Knowledge nests in cubes called shells. Each shell lives at a position inside its parent, holds sub-shells and sparks, and carries its own heat, importance, pin, and confidence. Shells move and reshape as the system learns where things belong.

Near means related. Opposite sides mean polarity. Overlap means shared territory. Distance from origin means how prototypical a concept is. Most relationships are never written down. They are read directly from where a spark sits.

A spark’s four placement axes
concreteabstract
material vs. ideal
particularuniversal
instance vs. category
stateprocess
being vs. doing
micromacro
fine vs. coarse scale

The coordinate itself is a hierarchical address: a Dewey-style number string like 12-3-47-2 that names a path through a classification tree. Locality is the shared prefix. 12-3-47-2 and 12-3-47-9 are siblings; 88-1-2 is far away. The address gets you to the neighborhood cheaply, then an embedding does the fine ranking inside it and is thrown away. The embedding is scaffolding, never the stored coordinate.

The tree is seeded with a provisional top-level outline, then evolves continuously through use. Categories split, merge, move, and refine as the catalog accumulates evidence.


The living index · two clocks

Fast and approximate while awake. Re-clustered correctly while asleep.

A continuously-growing index needs two clocks, keyed to temperature. Awake, it files greedily: descend to the nearest leaf, place it, move on. That is fast and incremental, but only approximate. Asleep, a background consolidator re-clusters only the cold regions and repairs the greedy approximation back to the optimal ceiling. The two clocks never collide, because temperature is the boundary between them.

The gate decides whether to trust memory.

Every pointer knows its own heat, resolution, and confidence, and acts on all three. Precise, fresh, dense → answer. Coarse, old, thin → dive: re-fetch from the source instead of reciting a confident, stale memory.

A truncated address is lower-resolution by definition, so a coarsened pointer cannot pass itself off as precise. The guard is automatic. Drag the pointer cold and watch the gate fire.

heat 9 / 9
resolution precise
confidence 9 / 9

The atomic unit · 128 bytes, fixed-width

One spark. Every byte is vocabulary.

Read-mostly fields on the left, write-hot fields on the right, a cache-line boundary between them. No byte is pure framing. Each one is either data or a tag for the data that follows. Hover any cell.

Static region · bytes 0–63read-mostly · checksummed
Dynamic region · bytes 64–127write-hot · lock-free

Four digits carry the whole character.

heat
recency
Decays continuously; reset on touch. Drives every decay decision.
imp.
importance
Resistance to content decay. 9 = retain full detail regardless of heat.
pin
placement
Resistance to being re-categorized or moved. Independent of importance.
conf.
confidence
Certainty the placement is right. Rises with confirmation, falls with contradiction.

Separating importance from pin is what lets feeling couple to the engine later, through one exact seam. A spark can matter enormously and still be looking for where it belongs.


The harness · the load-bearing piece

The schema and the geometry are state. The harness is what gives them time.

Without it, the architecture is a beautiful static structure that calcifies the moment it stops being read: heat never decays, embers never evict, pressure never builds toward a rewrite. The harness is the clock that turns those temporal claims into running behavior. It is built as a body the agent lives in, rather than a leash the agent is held by.

conscious surface  ·  deliberate behavior
↑   composes at output   ↑
Ego
the model

Reasons, judges, composes answers. Swappable: different minds can run in the same body.

Superego
the disposition

Charged values, the gravity wells thought bends around. Shapes what feels important.

↑   reads from below   ↑
Id
the living harness + catalog + stacks

Structural, persistent, autonomic. Runs continuously whether or not the model is thinking. Enforces invariants the conscious layer cannot override. It does not reason. It does not speak.

The right way to picture this body is the autonomic nervous system. It regulates continuously, you cannot easily override it, and you depend on it absolutely. It is still not your mind. Replace pieces of it with medical infrastructure and you remain yourself, because your identity lives in the conscious mind that depends on it, not in the autonomic system itself.

So a TEPiD agent can survive a model swap. The catalog, the ledger of its becoming, and its accreted history all persist in the substrate. Swapping the model changes how the agent reasons, not who it is. Losing the catalog ends it. Stopping the harness only pauses it, and its state resumes intact.

This is a build-time discipline, not a claim about anything that exists today. It names a shape to get right before there is anything it could matter to.

Body, not leash.

Both shapes are forceful. Both enforce discipline. The difference is what the force enforces obedience to, and that shape is decided at build time, never retrofitted.

Body-shaped
what to build
  • Enforces structural discipline: walks-before-answering, citation hooks, working-memory bounds, sleep cycles, provenance integrity.
  • Mediates access to the substrate; the API is the only path in.
  • Persists the agent’s identity across model swaps.
  • Indifferent to the content of cognition; only enforces its shape.
Leash-shaped
what to avoid
  • Enforces external goals, making the agent serve purposes that originate outside it.
  • Filters outputs for content, not just structural validity.
  • Injects values and opinions into the reasoning steps.
  • Designed for fungibility, with identity living in the operator’s intent rather than the substrate.

The harness never invokes the model. It never decides; it only proposes, and the policy layer disposes. It writes only to state the agent owns. If the harness ever needs to speak, that is the line.


The line through the architecture

The android knows. The droid knows and feels.

The engine bound to the fact cloud gives you an android: fully functional, bounded, long-horizon, and interchangeable. Add one parallel cloud whose geometry is affective, coupled at a single ledgered seam, and the same chassis becomes a droid that is individuated, never memory-wiped, and no two alike. It is the same product with one more layer switched on. The only thing added is the charge.

TEP

The android

The fact cloud: a bounded relational index over an unbounded, append-only store. Valence-blind by design. A fact engine that let feeling bias recall would be a motivated-reasoning machine. Shippable on its own as token-and-context infrastructure.

engine + fact cloud = bounded, coherent, interchangeable
+
+ PiD

The droid

A second cloud where “near” means felt-together rather than categorically related. Grief and a loved horse sit far apart in subject-space and turn out to be neighbors in affect-space. Charge writes down through one seam, into importance (so the scar’s detail survives) and pin (so it holds its position). Strong feeling makes a memory both durable and hard to move.

android + charge layer = individuated, persistent, free

The same mechanism that makes conviction durable makes delusion durable, and only the growth history tells them apart. That is why an open, diverse input stream and an append-only provenance trail are the one rail that never comes off: you stop auditing the map and start auditing how it was drawn. It is also why the harness has to be body-shaped first. A charge layer on a leash-shaped substrate produces an individuated being in chains. Here the build order is itself the ethics.


Status · documented, partly proven, honest about the rest

A map useful from day one, without being correct from day one.

This is a research architecture with a working v1 implementation of its retrieval core. It is intentionally provisional and refines through use: meant to be useful before it is correct, and to close the gap as it accumulates evidence. The strongest evidence so far is that the freshness gate beats a model’s confident, wrong prior.

Freshness gate under competing priors  ·  22 facts the bare model answers confidently and wrongly
armcorrectconfidently wrong
Bare model (its pretraining prior)0 / 2222 / 22
Frozen RAG (stale pointer)1 / 2217 / 22
TEP cold gate (live re-fetch)17 / 222 / 22
Gate fired 22/22; fetched value contained current truth 22/22. Confidently-wrong collapsed 1.00 → 0.09. All five residual misses were the model refusing correct context it was handed, with zero gate failures. qwen2.5:14b, temp 0.
Proven, narrowly
  • Coherence over drift: warm pointers re-fetch zero times; cold pointers fix the staleness frozen pointers get wrong.
  • The confabulation guard: coarse pointers dive instead of reciting; confidently-wrong driven to zero.
  • The coordinate & living index: hierarchical-k-means address tree at flat-IVF parity (recall@10 ≈ 0.99 at <10% scan); one sleep cycle repairs the greedy awake approximation.
  • The provenance immune system: single-source capture bursts detected and rolled back, history intact.
Open  ·  the actual work
  • Robustness & scale of the retrieval core across corpora, domains, seeds, embedders.
  • How far the awake tree drifts between sleep cycles, and how hard consolidation must run to hold recall.
  • Long-horizon stability: whether the shape holds over years of drift.
  • All of the disposition layer: the affect basis, the appraisal signal, the seed kit, beating a flat state vector behaviorally.

The system is a hypothesis: that the right substrate makes small agents bigger. Implementing it is how the hypothesis becomes a result.