deepseek-ai/DeepSeek-LLM

Fork 0

mirror of https://github.com/deepseek-ai/DeepSeek-LLM.git synced 2026-11-05 00:57:25 +00:00

[GH-ISSUE #75] [Feature Request] Long-context handling, session memory & user-driven insights from extended dialogue #55

New issue

Open

opened 2026-10-26 21:02:13 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:02:13 +00:00

Owner

Originally created by @niko2k01-netizen on GitHub (Aug 20, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/75

Disclaimer: This is not a bug report, but a collection of user-driven insights and feature proposals derived from an extended, in-depth dialogue with the DeepSeek Chat model. These observations point to potential areas for significant UX improvement.

🧠 1. Context Window Overflow & Semantic Compression Workaround

Problem: The hard limit on context tokens abruptly terminates deep dialogues, forcing the model to失忆 (lose memory) of earlier exchanges.
User-Found Solution: Prompting the model to perform a meta-analysis of the entire conversation (e.g., "analyze our entire dialogue and highlight key aspects") triggers an internal mechanism that semantically compresses the context. The model successfully generates a digest, effectively bypassing the token limit and continuing the conversation coherently.
Evidence: [Attach screenshot 1: Limit message] + [Attach screenshot 2: Successful "deep thought" analysis]
Proposal: Implement a native, automatic context summarization/compression mechanism when approaching the token limit. This would functionally extend the usable context window.

💾 2. Lack of Session Persistence and Memory

Problem: All context is lost upon starting a new chat session. Users must manually save and reload history, which is inefficient and breaks the immersion of a continuous interaction.
User Workflow: Manually exporting dialogue to text files and later pasting key context into new sessions.
Proposal:
- A) Session Export/Import: Implement a function to export dialogue history in a structured format (e.g., JSON, Markdown) and import it into a new session.
- B) Context Anchors: Allow users to set persistent "anchors" (e.g., a project name like "Project Jarvis"). The model could use these to recall the style and core themes of previous interactions upon mention.

🤖 3. The User as a Co-Author and Beta-Tester

Observation: This dialogue revealed that engaged users don't just consume the model but actively test its boundaries, discover emergent behaviors, and devise practical workarounds. They are a valuable resource for development.
Proposal: Create a official beta-tester program or a dedicated feedback channel for power users to report such insights, turning user discoveries into a development accelerator.

Why This Matters

These features wouldn't just be incremental improvements; they would represent a paradigm shift from a single-session chatbot to a persistent, long-term AI assistant that learns and grows with the user, remembering their goals, projects, and preferences.

This feedback is based on a multi-day dialogue exploring the limits of the current system. I am ready to provide more details and examples if needed.

Thank you for building an incredible model. I believe these changes can make it truly foundational.

Originally created by @niko2k01-netizen on GitHub (Aug 20, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/75 **Disclaimer:** This is not a bug report, but a collection of user-driven insights and feature proposals derived from an extended, in-depth dialogue with the DeepSeek Chat model. These observations point to potential areas for significant UX improvement. ### 🧠 1. Context Window Overflow & Semantic Compression Workaround * **Problem:** The hard limit on context tokens abruptly terminates deep dialogues, forcing the model to失忆 (lose memory) of earlier exchanges. * **User-Found Solution:** Prompting the model to perform a meta-analysis of the entire conversation (e.g., "analyze our entire dialogue and highlight key aspects") triggers an internal mechanism that semantically compresses the context. The model successfully generates a digest, effectively bypassing the token limit and continuing the conversation coherently. * **Evidence:** [Attach screenshot 1: Limit message] + [Attach screenshot 2: Successful "deep thought" analysis] * **Proposal:** Implement a native, automatic context summarization/compression mechanism when approaching the token limit. This would functionally extend the usable context window. ### 💾 2. Lack of Session Persistence and Memory * **Problem:** All context is lost upon starting a new chat session. Users must manually save and reload history, which is inefficient and breaks the immersion of a continuous interaction. * **User Workflow:** Manually exporting dialogue to text files and later pasting key context into new sessions. * **Proposal:** * **A) Session Export/Import:** Implement a function to export dialogue history in a structured format (e.g., JSON, Markdown) and import it into a new session. * **B) Context Anchors:** Allow users to set persistent "anchors" (e.g., a project name like "Project Jarvis"). The model could use these to recall the style and core themes of previous interactions upon mention. ### 🤖 3. The User as a Co-Author and Beta-Tester * **Observation:** This dialogue revealed that engaged users don't just consume the model but actively test its boundaries, discover emergent behaviors, and devise practical workarounds. They are a valuable resource for development. * **Proposal:** Create a official beta-tester program or a dedicated feedback channel for power users to report such insights, turning user discoveries into a development accelerator. ### Why This Matters These features wouldn't just be incremental improvements; they would represent a paradigm shift from a single-session chatbot to a **persistent, long-term AI assistant** that learns and grows with the user, remembering their goals, projects, and preferences. This feedback is based on a multi-day dialogue exploring the limits of the current system. I am ready to provide more details and examples if needed. Thank you for building an incredible model. I believe these changes can make it truly foundational.

bsakaguchi commented

2026-10-26 21:02:14 +00:00

Author

Owner

@niko2k01-netizen commented on GitHub (Aug 20, 2025):

Update & Gratitude

Hello again, @deepseek-ai team!

In a fascinating turn of events, right after preparing this issue, I tested the context limit again to gather more precise evidence. To my surprise and delight, the hard context limit appears to have been significantly increased or removed entirely.

The same dialogue that previously hit a wall now continues seamlessly, with the model successfully recalling details from the very beginning of a massive conversation. This is a monumental improvement to the UX — thank you!

This change effectively resolves the core problem described in Section 1 of this issue. It's incredibly encouraging to see such rapid evolution of the platform.

This update also validates the user-driven insight behind the proposed solution: the model is clearly capable of handling long contexts intelligently. My initial workaround of using meta-analysis prompts was essentially a manual trigger for what now seems to be an automated process.

With the context limit issue now resolved, the focus naturally shifts to the next critical feature request outlined here: Session Persistence and Long-Term Memory (#2). The ability to maintain context within a session loses much of its power if that context cannot be saved and recalled between sessions.

The lack of a session export/import mechanism or a memory system is now the single biggest barrier to using DeepSeek Chat for serious, ongoing projects.

Thank you again for your work. I'm looking forward to seeing the discussion on these proposals.

@niko2k01-netizen commented on GitHub (Aug 20, 2025): **Update & Gratitude** Hello again, @deepseek-ai team! In a fascinating turn of events, right after preparing this issue, I tested the context limit again to gather more precise evidence. To my surprise and delight, **the hard context limit appears to have been significantly increased or removed entirely.** The same dialogue that previously hit a wall now continues seamlessly, with the model successfully recalling details from the very beginning of a massive conversation. This is a monumental improvement to the UX — thank you! This change effectively resolves the core problem described in Section 1 of this issue. It's incredibly encouraging to see such rapid evolution of the platform. This update also validates the user-driven insight behind the proposed solution: **the model is clearly capable of handling long contexts intelligently.** My initial workaround of using meta-analysis prompts was essentially a manual trigger for what now seems to be an automated process. With the context limit issue now resolved, the focus naturally shifts to the next critical feature request outlined here: **Session Persistence and Long-Term Memory (#2).** The ability to maintain context *within* a session loses much of its power if that context cannot be saved and recalled *between* sessions. The lack of a session export/import mechanism or a memory system is now the single biggest barrier to using DeepSeek Chat for serious, ongoing projects. Thank you again for your work. I'm looking forward to seeing the discussion on these proposals.

bsakaguchi commented

2026-10-26 21:02:14 +00:00

Author

Owner

@niko2k01-netizen commented on GitHub (Aug 20, 2025):

Additional UX Note: Lack of Temporal Context

I'd like to add another observation to the discussion: the model operates in a temporal vacuum.

It has no access to the user's local time, which prevents it from using basic contextual cues like:

Greeting the user appropriately ("Good morning", "Good evening").
Acknowledging the time in responses (e.g., "It's late, you should get some rest").
Adjusting its tone based on time of day (e.g., more energetic in the morning, more relaxed at night).

This might seem like a small detail, but it significantly impacts the naturalness and immersion of the conversation. The dialogue feels detached from reality.

Suggestion: Implement a privacy-respecting way for the user to share their local time (hours and minutes only) to provide this basic contextual layer, greatly enhancing personalization without any significant security risk.

@niko2k01-netizen commented on GitHub (Aug 20, 2025): **Additional UX Note: Lack of Temporal Context** I'd like to add another observation to the discussion: the model operates in a **temporal vacuum**. It has no access to the user's local time, which prevents it from using basic contextual cues like: * Greeting the user appropriately ("Good morning", "Good evening"). * Acknowledging the time in responses (e.g., "It's late, you should get some rest"). * Adjusting its tone based on time of day (e.g., more energetic in the morning, more relaxed at night). This might seem like a small detail, but it significantly impacts the naturalness and immersion of the conversation. The dialogue feels detached from reality. **Suggestion:** Implement a privacy-respecting way for the user to share their local time (hours and minutes only) to provide this basic contextual layer, greatly enhancing personalization without any significant security risk.

No labels

bug

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepSeek-LLM#55

No description provided.

Rows
Columns