Continual In-Memory Map

Status: research scaffold with an imported prototype seed and an active C kernel experiment.

Renamed from "Continual In-Memory Neuromorphic Map" — the "neuromorphic" label is retired (decision D013, doc 06). The locality design principle (D004) is unchanged; the cinm_ code prefix is kept as a stable identifier.

This project captures a domain-neutral research concept for a small continual learning substrate. It is not an application, not a music system, and not a replacement for the existing ADG or Drum Engine work. Those domains may become case studies later. The first work is theoretical, mathematical, and systems oriented, with a small imported HDC playground preserved as an experiment seed.

Working name:

CIM: Continual In-Memory Map

(The code keeps the cinm_ prefix as a stable identifier.)

Core thesis:

event stream
  -> sparse context addressing
  -> local memory-cell activation
  -> bounded local update
  -> immediate readout
  -> periodic consolidation

The experiment asks whether a small local memory field can learn continuously from events without global retraining, while remaining inspectable, bounded, and useful immediately after each event.

Current terminology for the human/evidence-gated adaptation line:

interactive continual preference learning
  with reversible self-adaptation

Motivation

The starting discomfort is that modern neural networks make training the central operation. They often require large global optimization passes, expensive dense compute, and poor continual learning behavior. This project explores a different regime:

• continual learning as the normal case
• memory and computation kept together
• sparse activation instead of full-model activation
• local plasticity instead of global retraining
• append-only learning evidence instead of opaque weight churn
• bounded state change, decay, conflict handling, and consolidation

The goal is not to beat transformers at their own task. The goal is to define and test a different learning substrate.

Reading Order

1. docs/00-conversation-summary.md
2. docs/01-theory.md
3. docs/02-system-architecture.md
4. docs/03-research-program.md
5. docs/04-language-lines.md
6. docs/05-decisions.md
7. docs/06-open-questions.md
8. docs/07-future-domain-bridge.md
9. docs/08-imported-taste-hdc-seed.md
10. docs/09-market-and-research-landscape.md
11. docs/10-neurosymbolic-transfer.md
12. docs/11-c-cpu-simd-hdc.md
13. docs/12-memory-model.md
14. docs/13-implementation-backlog-g-k.md
15. docs/14-tail-latency-hedging.md
16. docs/15-drum-vertical-architecture.md
17. docs/16-mvp-milestones.md
18. docs/17-implementation-backlog-l-q.md
19. docs/18-evaluation-and-baseline-plan.md
20. docs/19-claim-boundaries-and-authority.md
21. docs/20-risk-register.md
22. docs/21-glossary.md
23. docs/22-state-primacy-refactor.md
24. docs/FORMAL_PROOF.md

Deferred / future mapping (not in current RAM-only scope, decision D013):

• docs/future/ssd-store-linux-layer.md — durable SSD store + Linux deep dive.

Layout

• docs/ records the theory, decisions, and research program.
• docs/15–docs/21 define the first applied case study — the drum-first vertical (D014): architecture, MVP milestones, backlog L–Q, evaluation, claim boundaries, risks, and glossary.
• docs/22 is the state-primacy refactor plan (D018): the staged R0–R5 work that brings the C kernel in line with "the live map is primary; replay is a within-epoch verification property." It lands before the drum vertical consumes the substrate.
• AGENTS.md is the entrypoint for an agent implementing the drum vertical: file map, phase plan, reuse map, hard rules, and verification.
• docs/isabelle/ contains the minimal Isabelle/HOL proof lane for abstract replay, snapshot-tail equivalence, and checked replay rejection.
• experiments/taste-hdc/ preserves the imported CPU + RAM HDC taste-memory seed from /home/dev/taste-hdc.
• experiments/c-kernel/ contains the first C substrate: SoA synaptic map, pairwise update, sigma A/B harness, reversible transaction check, and an in-RAM event-log replay/recovery check, plus G-K evidence, learning, memory, and HDC experiment gates. Strictly in-memory — no disk tier (decision D013).

The imported seed is intentionally small and exploratory. It is not the final CINM implementation contract.

Non-Goals

• No production implementation claim.
• No ADG, Drum Engine, PC4MS, or music-specific dependency in the core concept.
• The drum vertical (D014, docs 15–21) is a deliberate applied case study, fenced by claim boundaries (doc 19); the core concept stays domain-neutral (D001).
• No claim that this is biologically faithful neuroscience.
• No claim that this replaces backpropagation for large-scale representation learning.
• No production claim, benchmark claim, or quality claim.

First Artifact Target

The first durable artifact should remain a clean research note and experiment plan. The imported prototype seed can inform this work, but it should not silently become the theory.

theory
  + state model
  + local update rules
  + consolidation rules
  + toy task suite
  + language/backend split
  + explicit claim boundaries