mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-28 11:40:07 +00:00
Update README.md
This commit is contained in:
PSBigBig
GitHub
parent
38e10c7f4a
commit
c6db6bb964
1 changed files with 44 additions and 107 deletions
|
|
@ -1,125 +1,62 @@
|
|||
# Memory & Long-Context — Global Fix Map
|
||||
|
||||
Keep threads coherent across long windows and session restarts.
|
||||
Detect and repair entropy melt, boundary drift, and state desync.
|
||||
Detect and repair entropy melt, boundary drift, state forks, and ghost contexts.
|
||||
|
||||
---
|
||||
|
||||
## What this page is
|
||||
- A compact checklist for long contexts and multi-session memory
|
||||
- Copyable guards to stop drift and collapse before they spread
|
||||
- How to measure stability with ΔS and λ_observe
|
||||
- A compact checklist for long contexts and multi-session memory.
|
||||
- Copyable guardrails to stop drift and collapse before they spread.
|
||||
- Concrete measures with ΔS and λ_observe to verify stability.
|
||||
|
||||
---
|
||||
|
||||
## When to use
|
||||
- Dialogs grow past 50k to 100k tokens and answers degrade
|
||||
- Facts flip after tab refresh or model switch
|
||||
- Citations look right yet reasoning goes flat or chaotic
|
||||
- OCR transcripts look fine but capitalization and spacing drift
|
||||
- Multi day support threads lose task state or rewrite history
|
||||
|
||||
## Open these first
|
||||
- Session continuity and state fences: [Memory Coherence](https://github.com/onestardao/WFGY/blob/main/ProblemMap/memory-coherence.md)
|
||||
- Long window drift and attention melt: [Entropy Collapse](https://github.com/onestardao/WFGY/blob/main/ProblemMap/entropy-collapse.md)
|
||||
- Long reasoning chain drift: [Context Drift](https://github.com/onestardao/WFGY/blob/main/ProblemMap/context-drift.md)
|
||||
- Cross tab and cache hazards: [Memory Desync Pattern](https://github.com/onestardao/WFGY/blob/main/ProblemMap/patterns/pattern_memory_desync.md)
|
||||
- Trace schema and audit trail: [Retrieval Traceability](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md)
|
||||
- Snippet and citation schema: [Data Contracts](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md)
|
||||
- Chunk stability at joins: [Chunking Checklist](https://github.com/onestardao/WFGY/blob/main/ProblemMap/chunking-checklist.md)
|
||||
- OCR quality and normalization: [OCR Parsing Checklist](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ocr-parsing-checklist.md)
|
||||
- Dialogs grow past 50k–100k tokens and answers degrade.
|
||||
- Facts flip after tab refresh or model switch.
|
||||
- Citations look right but reasoning goes flat or chaotic.
|
||||
- OCR transcripts look fine but capitalization and spacing drift.
|
||||
- Multi-day support threads lose task state or rewrite history.
|
||||
|
||||
---
|
||||
|
||||
## Common failure patterns
|
||||
- **Entropy melt** attention variance climbs with length and the model smooths meaning
|
||||
- **Boundary leak** chunks merge across section joins and citations shift by a few lines
|
||||
- **State fork** two tabs or agents hold different memory revisions and answers flip
|
||||
- **Ghost context** stale buffers linger after role or persona change and contaminate steps
|
||||
- **OCR jitter** mixed spacing or width variants create false token differences
|
||||
## Quick routes to per-page guides
|
||||
|
||||
- Memory fences and continuity
|
||||
→ [memory-coherence.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/memory-coherence.md)
|
||||
|
||||
- Attention melt in long windows
|
||||
→ [entropy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/entropy-collapse.md)
|
||||
|
||||
- Long reasoning drift
|
||||
→ [context-drift.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/context-drift.md)
|
||||
|
||||
- Cross-tab and cache hazards
|
||||
→ [pattern_memory_desync.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/pattern_memory_desync.md)
|
||||
|
||||
- Stale buffers and persona residue
|
||||
→ [ghost-context.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/ghost-context.md)
|
||||
|
||||
- Divergent memory forks
|
||||
→ [state-fork.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/state-fork.md)
|
||||
|
||||
- OCR-specific noise
|
||||
→ [ocr-parsing-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/ocr-parsing-checklist.md), [ocr-jitter.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/ocr-jitter.md)
|
||||
|
||||
- Traceability and audit trail
|
||||
→ [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/retrieval-traceability.md), [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/data-contracts.md)
|
||||
|
||||
- Chunk stability at joins
|
||||
→ [chunking-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/MemoryLongContext/chunking-checklist.md)
|
||||
|
||||
---
|
||||
|
||||
## Fix in 60 seconds
|
||||
1) **Stamp and fence state**
|
||||
- At turn start set `mem_rev`, `mem_hash`, `task_id`
|
||||
- Forbid writes if client stamps do not match the server record
|
||||
|
||||
2) **Shard the window**
|
||||
- Assemble prompts as `{system | task | constraints | snippets | answer}`
|
||||
- Split snippets by section and forbid cross section reuse
|
||||
|
||||
3) **Normalize inputs**
|
||||
- Unicode NFC, strip zero width, unify full and half width
|
||||
- Drop OCR lines below confidence threshold
|
||||
|
||||
4) **Stabilize attention**
|
||||
- Apply BBAM to clamp variance
|
||||
- If collapse detected use BBCR to bridge and re anchor
|
||||
|
||||
5) **Probe the joins**
|
||||
- Measure ΔS across adjacent chunks and keep each join ≤ 0.50
|
||||
- Plot ΔS(question, retrieved) vs k and expect a downward curve after the fix
|
||||
|
||||
6) **Trace or stop**
|
||||
- Require cite then answer
|
||||
- If a claim has no snippet id stop and ask for the exact citation
|
||||
|
||||
---
|
||||
|
||||
## Copy paste prompt
|
||||
```
|
||||
|
||||
You have TXT OS and the WFGY Problem Map.
|
||||
|
||||
Goal
|
||||
Stabilize memory across long windows and across sessions without losing traceability.
|
||||
|
||||
Protocol
|
||||
|
||||
1. Print {mem\_rev, mem\_hash, task\_id}. If missing set defaults and echo them.
|
||||
2. Build a Snippet Table with columns {section\_id | start\_line | end\_line | citation}.
|
||||
3. Guardrails
|
||||
|
||||
* cite then answer
|
||||
* forbid cross section reuse
|
||||
* if a claim lacks a snippet id stop and request it
|
||||
4. Collapse control
|
||||
|
||||
* if attention variance rises apply BBAM
|
||||
* if logic stalls apply BBCR and show the bridge node
|
||||
5. Metrics
|
||||
|
||||
* report ΔS(question, retrieved)
|
||||
* report ΔS across each join
|
||||
* report λ\_observe at retrieval, assembly, reasoning
|
||||
|
||||
Input
|
||||
|
||||
* question
|
||||
* snippets with ids and line ranges
|
||||
* previous {mem\_rev, mem\_hash, task\_id} if any
|
||||
|
||||
Output
|
||||
|
||||
* header {mem\_rev, mem\_hash, task\_id}
|
||||
* Snippet Table
|
||||
* Bridge Check
|
||||
* Final Answer with inline citations
|
||||
* ΔS and λ states
|
||||
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Minimal checklist
|
||||
- State stamped with `mem_rev` and `mem_hash` at every turn
|
||||
- Prompt schema locked and section fences enforced
|
||||
- Unicode normalized and OCR noise gated
|
||||
- BBAM enabled and BBCR available on collapse
|
||||
- ΔS at each join ≤ 0.50 and overall ΔS(question, retrieved) ≤ 0.45
|
||||
- Cite then answer and no orphan claims
|
||||
|
||||
## Acceptance targets
|
||||
- Retrieval coverage ≥ 0.70 to the intended section
|
||||
- ΔS(question, retrieved) ≤ **0.45** and joins ≤ **0.50**
|
||||
- λ remains **convergent** across three paraphrases
|
||||
- No state fork across tabs or agents for the same `task_id`
|
||||
- No state fork across tabs or agents for the same `task_id`
|
||||
|
||||
---
|
||||
|
||||
|
|
@ -127,7 +64,7 @@ Output
|
|||
|
||||
| Tool | Link | 3-Step Setup |
|
||||
|------|------|--------------|
|
||||
| **WFGY 1.0 PDF** | [Engine Paper](https://github.com/onestardao/WFGY/blob/main/I_am_not_lizardman/WFGY_All_Principles_Return_to_One_v1.0_PSBigBig_Public.pdf) | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + \<your question>” |
|
||||
| **WFGY 1.0 PDF** | [Engine Paper](https://github.com/onestardao/WFGY/blob/main/I_am_not_lizardman/WFGY_All_Principles_Return_to_One_v1.0_PSBigBig_Public.pdf) | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
|
||||
| **TXT OS (plain-text OS)** | [TXTOS.txt](https://github.com/onestardao/WFGY/blob/main/OS/TXTOS.txt) | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
|
||||
|
||||
---
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue