This release is a pre-release and may not be stable for production use.
plato
A memory that any LLM can use. plato sits between your app and the model you already pay for (by API key - GPT, Claude, Grok, Kimi, or your own server) as an OpenAI-compatible proxy. A small trained controller (Qwen3.5-2B) operates a Differentiable Neural Computer (DNC) memory - a fixed grid of slots with learned write, read, link and time mechanics - remembers what happened in past conversations, renders what it read as labeled text into your request, and maintains itself while idle (merging duplicates, resolving contradictions toward newer facts, pruning stale entries). You pay for the size of the memory, in slots - never for your model's tokens.
Documents
| Doc | What it is |
|---|---|
| docs/mvp-design.md | The design of record - rev 19, LOCKED 2026-08-31 after nine adversarial review rounds. Architecture, models, training, evaluation gates, serving, pricing, milestones, risks, and all owner rulings (Q1-Q28). Starts with a plain-language one-page overview. |
| docs/m0-implementation.md | The M0 implementation plan - rev 14, LOCKED. The first milestone: repo scaffold, PyTorch DNC port with golden parity against the DeepMind reference, training smoke tests, and the throughput benchmarks that replace the design's cost estimates. |
| docs/m0-execution-plan.md | The M0 execution plan - rev 3, DRAFT. The M0 plan broken into PR-sized tasks (S1-S9) with named files, tests, mutants, fixtures and exit-gate traceability; implementer decisions and doc-vs-doc discrepancies flagged for errata; adversarial review log (Codex + Grok). Subordinate to the M0 plan. |
| docs/architecture-explainer.md | Plain-language companion with ASCII diagrams: the serving path, one controller turn, the anatomy of a memory slot, and the training mechanism (three-tier signal ladder, two-phase BPTT). Non-normative. |
| docs/xacp-trust-boundary-note.md | Investigation note (non-normative, post-MVP): embedding XACP-style selective sharing - what knowledge to share and with whom - via write-time layering, one memory per disclosure layer. |
Where a document disagrees with docs/mvp-design.md, the design doc wins.
Status
M0 COMPLETE (2026-09-03, under the rev-16 erratum). The design is at rev 21 - the environment pivot (Q29-Q32): three surfaces over one op set, reader-free-at-scale training. Next: the M1 execution-plan delta pass, then M1. Report: docs/m0-report.md.
License
1 maintainer
Metadata
1 maintainer
Release files for plato-ai 0.1.0.dev0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| plato_ai-0.1.0.dev0.tar.gz | 1.6 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| plato_ai-0.1.0.dev0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 2.7 MB
Release files / plato_ai-0.1.0.dev0.tar.gz
| Download URL | plato_ai-0.1.0.dev0.tar.gz |
|---|---|
| Size | 1.6 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
a37e36f7839ab0ebe6bc2a9fa6216150148ced3fc29a3b11c7332756a7af2c9e
|
|
BLAKE2b-256 checksum How to use checksums |
2e1fb74bbee66f5a1d31cadcc451a6154c1e76bfe3bee20c21d2429efbd8c4fc
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.14
|
Release files / plato_ai-0.1.0.dev0-py3-none-any.whl
| Download URL | plato_ai-0.1.0.dev0-py3-none-any.whl |
|---|---|
| Size | 1.1 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
17c31819cc1473454e0a2efcdba9aa8e794212592effd8cae283f84f4e1d6e44
|
|
BLAKE2b-256 checksum How to use checksums |
4056de01b7aa7f7c35b4aaf4e3c39678f53f933b0cbd3f358c8dd84efcfad94b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.14
|
