dopetalk does not endorse any advertised product nor does it accept any liability for it's use or misuse


Our Discord Notification Server invitation link is https://discord.gg/jB2qmRrxyD
« previous next »
Pages: 1   Go Down

Author Topic: Project "Influence" Master -- is an "organic" memory mechanisn (LLM2) viable ?  (Read 126 times)

Online smfadmin (OP)

  • SMF (internal) Site
  • Owner
  • Sr. Member
  • *****
  • Administrator
  • *****
  • Admin
  • ****
  • Join Date: Dec 2014
  • Location: Management
  • Posts: 694
  • Reputation Power: 0
  • smfadmin has hidden their reputation power
  • Last Login:Today at 05:18:35 PM
  • Supplied Install Member
Project "Influence" Master -- is an "organic" memory mechanisn (LLM2) viable ?
« on: October 27, 2026, 12:45:28 AM »
📎 Click on me to view all my attachments
ListAttBBC  |  v7.3.0  |  2026-10-26  |  Andrew.human & Claude.ai  |  live 17:45:12
Project: 3 posts · master: Project "Influence" Master -- is an "organic" memory mechanisn (LLM2) viable ? (topic 8078) (chosen by hand) · 2026-10-11 to 2026-10-26 (15 days)
#FileSizeDownloadsInfoDL
00NEXUS.md2.3 KB2597ℹ️⬇️
01Experiment_001_checkpoint.md7.1 KB3ℹ️⬇️
02Qwen on 'How it would influence the REPLY from an LLM is a way that typical MEMORY Mecahnisms do, but without changing the QUERY'.pdf  [native]  [pdf in chunks]393.8 KB8ℹ️⬇️
03MEMSTORE_CLONE_CHECKPOINT_2026-10-27_2.md8.4 KB2ℹ️⬇️
04HOTWORDS_MANUAL.md13.4 KB3ℹ️⬇️
05ContextDef05b_v1.0.0.zip8 KB3ℹ️⬇️
06CONTEXTDEF05B_MANUAL.md5.6 KB3ℹ️⬇️
07Project 'Influence' homework and prep. -- chapter 1 CLOSED.pdf  [native]  [pdf in chunks]1.2 MB8ℹ️⬇️
08Experiment_001_checkpoint_2.md14.5 KB3ℹ️⬇️
09Experiment_001_checkpoint_3.md12.1 KB3ℹ️⬇️
10Project Influence Chapter 2A.pdf  [native]  [pdf in chunks]629.8 KB8ℹ️⬇️
11Project Influence Chapter 2B.pdf  [native]  [pdf in chunks]1.6 MB8ℹ️⬇️
12Experiment_001_checkpoint_5.md9.9 KB2ℹ️⬇️
13Experiment_001_checkpoint_4.md9.6 KB2ℹ️⬇️
14Experiment_001_checkpoint_6.md16.1 KB2ℹ️⬇️
15NATURAL_RECORDER_MANUAL.md8.2 KB2ℹ️⬇️
16Memstore_Clone_v1.0.16.zip50 KB2ℹ️⬇️
17MEMSTORE_MANUAL.md6.9 KB2ℹ️⬇️
18QWEN_MEMCHAT_MANUAL.md5.7 KB2ℹ️⬇️
19QWEN_MEMPROXY_MANUAL.md16.4 KB2ℹ️⬇️
20QWEN_PROPOSE_MANUAL.md6.3 KB2ℹ️⬇️
21Qwen25_05B.zip10.4 KB2ℹ️⬇️
22Experiment_001_checkpoint_7.md20 KB2ℹ️⬇️
23Project Influence Chapter 3 CLOSED.pdf  [native]  [pdf in chunks]3.1 MB9ℹ️⬇️
24NaturalRecorder_v1.2.1.zip16.8 KB3ℹ️⬇️
25Exp000_queries_v1.0.0.zip2.1 KB3ℹ️⬇️
26instagram-1793269250064.jpg159.8 KB7ℹ️⬇️
27Project UNDER THE Influence LOL -- Chapter 4 get trashed and dies.pdf  [native]  [pdf in chunks]2 MB5ℹ️⬇️
28Project Influence Chapter 4 -turns to trash, lol, sorry.pdf  [native]  [pdf in chunks]1.9 MB5ℹ️⬇️
29Experiment_001_checkpoint_8.md21.7 KB1ℹ️⬇️
30Experiment_001_checkpoint_9.md11.3 KB1ℹ️⬇️
31Hotwords_v2.2.0.zip28.7 KB1ℹ️⬇️
32Experiment_001_checkpoint_10.md9.5 KB1ℹ️⬇️
33Project Influence Chapter 5 CLOSES.pdf  [native]  [pdf in chunks]1.2 MB5ℹ️⬇️
34standard_prompt.txt3 KB4ℹ️⬇️
35memtok_candidates.txt466 B4ℹ️⬇️
36ChatGPT_reply.txt2.5 KB4ℹ️⬇️
37El Anthroppo Claude 'E' Shennanegans.pdf  [native]  [pdf in chunks]167.8 KB2ℹ️⬇️
38Logits in LLMs - Chapter 6 'UNDER THE AFFLUENT 'lol the CLOSER.pdf  [native]  [pdf in chunks]767.5 KB2ℹ️⬇️
39Influence_Chapter_6_checkpoint.md9.9 KB1ℹ️⬇️
40Which Loose Larry Left their Logits in 'le' (c'est Francais) LLM, hmmm.pdf  [native]  [pdf in chunks]124.9 KB2ℹ️⬇️

This is an experiment to see if/how AI could benefit if "influenced" by a datastore of it's, initially, QUERY/REPLY historical data - using a second LLM (inferred reply not used) to INFLUENCE LLM1's REPLY, by selective weighting changes.

Steering a LLM Under the Influence

Logits are the raw, unnormalized predictions (outputs) generated by a neural network before they are converted into probabilities in the final layer

The AI guy:

  • No recorder attached: the next conversation needs NaturalRecorder_v1.3.0.zip. Only 1.2.1 reached me this time, so check which one you have.
  • Embedding server: it runs in the foreground from embed-n-serve.sh, so it will not come back after the NUC shuts down. Start it again before any pseudo --hybrid run. If it is down, hybrid falls back to word overlap with a warning.
  • Store stage: the checkpoint records hotwords_store_stage.py as on, per checkpoint 9. I did not recheck it, so it stays on until you switch it off.
  • Test store: /root/memstore/hwtest holds the three test arrays with their vectors, ready for the next test.

The next chapter starts from the checkpoint's open items. Enjoy the rest of the House set.

Me: I think that we should forget all of my previous plans and go bananas on logits - so tell me what they can and cannot do and why we need look no further ?

The AI guy: What logits can do

  • Move any token's chance up or down by an exact amount. Adding to a logit multiplies that token's odds by a fixed factor, and negative infinity bans a token outright.
  • Act at every step of generation, so one standing bias list works as the always-on BASE RULES layer.
  • Leave the model untouched. No retraining, and you can switch it on and off per run.
  • Make every change auditable. Logging the logits before and after your edit gives you the traceable, separate stage you wanted, and the LLM1-normal versus LLM1-influenced comparison in Experiment 001 falls out for free.
  • Carry influence forward indirectly. The token that gets chosen becomes context for everything after it.

What logits cannot do

  • Work on words. The model scores tokens, and a word or mini-phrase may be several tokens, with the leading-space variant being a different token again. A boost on a multi-token word only bites on its first piece, so the store needs to map each mem-tok to its token sequences and keep boosting as the sequence proceeds.
  • Change what the model understands. Logits sit after all the attention and reasoning, so they alter what gets said and not how the context is represented inside.
  • Know whether a token fits. The boost is the same size wherever the token could appear, so a heavy boost on something the model finds implausible gives fluent nonsense or a loop, not a corrected belief.
  • Handle meaning. Synonyms and paraphrases must be listed one by one.
  • Remember. The model is stateless, so the memory lives in your store.
  • Work everywhere. Hosted models expose little or nothing of the logits, but your local Qwen under llama.cpp gives you all of them.
  • Be compared across models. Two models need the same tokenizer for their pools to line up, and as I said earlier, average log-probabilities rather than raw logits, because raw logits carry an arbitrary offset.

On "look no further"

It's the one chokepoint everything passes through, and the only place with exact, additive, traceable control, so it earns the lead role. But I'd disagree with "no further" on two counts.

  1. On 2026-10-27 you wanted finer control of the token-attention relationship, and numeric poking at internal context. That sits upstream of the logits, in the hidden states, and logits can't reach it. Logits decide what gets said, and the internal layers decide what the model is thinking.

  2. You also said messing with reply tokens is risky and moved Experiment 002 to prefill injection. Boosting logits on reply tokens is that same risky move, done more precisely. It's fine as an experiment, but it reopens the question you closed.

The clean split would be logits as the actuator for 001 and the mem-tok store, with prefill and internal-state work kept as the separate track for 002.

Me: 001=LLM-Uno and 002=LLM-Duo, n'est pas ?

The AI guy: Non, not quite. Both experiments use both models. LLM1 is the one that writes the reply, and LLM2 is the influencer fed by the attention and RAG store.

What separates the two experiments is what LLM2 is allowed to do to LLM1:

  • 001: LLM2 only reweights words already among LLM1's own top-50 candidates, and adds nothing new.
  • 002: LLM2 also adds words LLM1 wasn't considering, with artificially heavier weights, and you later chose to do that by injecting tokens at the start of the reply, before any reply token is selected.

So if you want the Uno/Duo names, they fit better as labels for the experiments than for the models. "001 = reweight only" and "002 = reweight plus injection" would be the accurate version.

Me: TY, common sense prevails; additional words may enrichen context but will remain foreign to the logits

The AI guy: Mostly right, with one correction. Words injected into the context reach the logits only through the model's own computation: attention reads them, the hidden state shifts, and the scores follow. Nothing hands them a score directly, so the model keeps the final say. That's the safe property of the context route, and it's why prefill injection is gentler than boosting logits.

Where I'd soften "foreign": models have a strong habit of copying words that appear in their context, so an injected word often gets a real lift in its own logit, and more so when it's weighty or repeated. It isn't foreign, it's just not guaranteed. A word the model finds implausible in that position can still sit low however richly it's in the context.

That gives you two levers with different guarantees. Context injection is a suggestion that the model may take or ignore, while a logit boost is a command it can't refuse. For the mem-tok idea you could use the first to enrich and the second sparingly to force.



📝 Inline Markdownv4.6.2 · 2026-10-28

The Principles of Operation include (but are not limited to) the follwing:

  1. 'Influence' supports the three main memory emulations: episodic (internal), semantic (external) and procedural (operational)
  1. The "currency" of 'Influence' are mem-toks (aka 'memory tokens') and are keyed by either a whole word, a word pair or a mini-phrase, whereby influence is exerted by either injecting new or existing RELATIONAL (to LLM1's Top 20-50) tokens, chosen by LLM2's Attention+RAG, with the assigned significance of a high net worth, to "force" LLM1 to pay Attention !
  1. We may choose to exploit the internal vocabulary and  see if it's worth associating the RELATIONAL tokens with synonyms.

--

  1. And we must not overlook Logits and how we can best manipulate them. 



AKA (under the) Influence, Chapter 4 closes the door

The nominated embedding server was supplied by https://huggingface.co/Felladrin/gguf-Q8_0-all-MiniLM-L6-v2/tree/main with the command "wget https://huggingface.co/Felladrin/gguf-Q8_0-all-MiniLM-L6-v2/resolve/main/all-minilm-l6-v2-q8_0.gguf"

« Last Edit: Today at 07:28:46 AM by smfadmin »
friendly
0
funny
0
informative
0
agree
0
disagree
0
like
0
dislike
0
No reactions
No reactions
No reactions
No reactions
No reactions
No reactions
No reactions
measure twice, cut once

Online smfadmin (OP)

  • SMF (internal) Site
  • Owner
  • Sr. Member
  • *****
  • Administrator
  • *****
  • Admin
  • ****
  • Join Date: Dec 2014
  • Location: Management
  • Posts: 694
  • Reputation Power: 0
  • smfadmin has hidden their reputation power
  • Last Login:Today at 05:18:35 PM
  • Supplied Install Member
Re: Project "Influence" Master -- is an "organic" memory mechanisn (LLM2) viable ?
« Reply #1 on: Yesterday at 06:23:11 PM »
Topic bumped
friendly
0
funny
0
informative
0
agree
0
disagree
0
like
0
dislike
0
No reactions
No reactions
No reactions
No reactions
No reactions
No reactions
No reactions
measure twice, cut once
Pages: 1   Go Up
« previous next »
Tags:
 

dopetalk does not endorse any advertised product nor does it accept any liability for it's use or misuse




TERMS AND CONDITIONS

In no event will d&u or any person involved in creating, producing, or distributing site information be liable for any direct, indirect, incidental, punitive, special or consequential damages arising out of the use of or inability to use d&u. You agree to indemnify and hold harmless d&u, its domain founders, sponsors, maintainers, server administrators, volunteers and contributors from and against all liability, claims, damages, costs and expenses, including legal fees, that arise directly or indirectly from the use of any part of the d&u site.

TO USE THIS WEBSITE YOU MUST AGREE TO THE TERMS AND CONDITIONS ABOVE

Founded December 2014
SimplePortal 2.3.6 © 2008-2014, SimplePortal