Asosiy tarkibga o'tish

​
Workspace Memory v2 (oflayn): tadqiqot qaydlari

Maqsad: Clawd uslubidagi ish joyi (agents.defaults.workspace, standart ~/.openclaw/workspace), bu yerda “xotira” har bir kun uchun bitta Markdown faylida (memory/YYYY-MM-DD.md) hamda bir nechta barqaror fayllarda (masalan, memory.md, SOUL.md) saqlanadi. Ushbu hujjat Markdown’ni ko‘rib chiqiladigan, kanonik manba-haqiqat sifatida saqlab qoladigan, biroq hosila indeks orqali tuzilmali eslash (qidiruv, entitet xulosalari, ishonch yangilanishlari) ni qo‘shadigan oflayn-birinchi xotira arxitekturasini taklif qiladi.

​
Nega o‘zgartirish kerak?

Joriy sozlama (kuniga bitta fayl) quyidagilar uchun juda yaxshi:
  • “faqat qo‘shib boriladigan” jurnal yuritish
  • inson tomonidan tahrirlash
  • git asosidagi barqarorlik + audit qilinuvchanlik
  • kam ishqalanishli qayd etish (“shunchaki yozib qo‘yish”)
U quyidagilar uchun zaif:
  • yuqori darajadagi qayta chaqirish (“X haqida nimaga kelishgandik?”, “oxirgi marta Y ni qachon sinab ko‘rdik?”)
  • ko‘p fayllarni qayta o‘qimasdan entitetga yo‘naltirilgan javoblar (“Alice / The Castle / warelay haqida aytib ber”)
  • fikr/afzalliklar barqarorligi (va u o‘zgarganda dalillar)
  • vaqt cheklovlari (“2025-yil noyabrida nima to‘g‘ri edi?”) va ziddiyatlarni hal qilish

​
Dizayn maqsadlari

  • Oflayn: tarmoqsiz ishlaydi; noutbukda/Castle’da ishga tushadi; bulutga bog‘liqlik yo‘q.
  • Tushuntiriladigan: qaytarib olingan elementlar manbaga bog‘lanishi (fayl + joylashuv) va inferensiyadan ajratilishi kerak.
  • Kam marosim: kundalik yozuvlar Markdown bo‘lib qoladi, og‘ir sxema ishlari yo‘q.
  • Bosqichma-bosqich: v1 faqat FTS bilan ham foydali; semantik/vektor va graf funksiyalari ixtiyoriy yangilanishlardir.
  • Agentga qulay: “token byudjeti doirasida eslab qolish”ni osonlashtiradi (faktlarning kichik to‘plamlarini qaytaradi).

​
North star model (Hindsight × Letta)

Birlashtiriladigan ikki qism:
  1. Letta/MemGPT-style control loop
  • kichik “asosiy” qismni doimo kontekstda saqlash (persona + foydalanuvchi haqidagi muhim faktlar)
  • qolgan hamma narsa kontekstdan tashqarida bo‘ladi va vositalar orqali chaqirib olinadi
  • xotiraga yozish aniq tool chaqiruvlari (append/replace/insert) orqali amalga oshiriladi, saqlanadi va keyingi navbatda qayta kiritiladi
  1. Hindsight-style memory substrate
  • kuzatilgan, ishonilgan va umumlashtirilgan ma’lumotlarni alohida ajratish
  • support retain/recall/reflect
  • confidence-bearing opinions that can evolve with evidence
  • entity-aware retrieval + temporal queries (even without full knowledge graphs)

​
Proposed architecture (Markdown source-of-truth + derived index)

​
Canonical store (git-friendly)

Keep ~/.openclaw/workspace as canonical human-readable memory. Suggested workspace layout: 
~/.openclaw/workspace/
  memory.md                    # small: durable facts + preferences (core-ish)
  memory/
    YYYY-MM-DD.md              # daily log (append; narrative)
  bank/                        # “typed” memory pages (stable, reviewable)
    world.md                   # objective facts about the world
    experience.md              # what the agent did (first-person)
    opinions.md                # subjective prefs/judgments + confidence + evidence pointers
    entities/
      Peter.md
      The-Castle.md
      warelay.md
      ...
Notes:
  • Daily log stays daily log. No need to turn it into JSON.
  • The bank/ files are curated, produced by reflection jobs, and can still be edited by hand.
  • memory.md remains “small + core-ish”: the things you want Clawd to see every session.

​
Derived store (machine recall)

Add a derived index under the workspace (not necessarily git tracked): 
~/.openclaw/workspace/.memory/index.sqlite
Back it with:
  • SQLite schema for facts + entity links + opinion metadata
  • SQLite FTS5 for lexical recall (fast, tiny, offline)
  • Loglar JSONL formatida (har bir qatorda bitta JSON obyekt).
Semantik eslab qolish uchun ixtiyoriy embeddinglar jadvali (hali oflayn).

​
Retain / Recall / Reflect (operational loop)

​
Retain: normalize daily logs into “facts”

Hindsight’s key insight that matters here: store narrative, self-contained facts, not tiny snippets. Practical rule for memory/YYYY-MM-DD.md:
  • at end of day (or during), add a ## Retain section with 2–5 bullets that are:
    • narrative (cross-turn context preserved)
    • self-contained (standalone makes sense later)
    • tagged with type + entity mentions
Example: 
## Retain
- W @Peter: Currently in Marrakech (Nov 27–Dec 1, 2025) for Andy’s birthday.
- B @warelay: I fixed the Baileys WS crash by wrapping connection.update handlers in try/catch (see memory/2025-11-27.md).
- O(c=0.95) @Peter: Prefers concise replies (<1500 chars) on WhatsApp; long content goes into files.
Minimal parsing:
  • Type prefix: W (world), B (experience/biographical), O (opinion), S (observation/summary; usually generated)
  • Entities: @Peter, @warelay, etc (slugs map to bank/entities/*.md)
  • Opinion confidence: O(c=0.0..1.0) optional
Indeks har doim Markdown’dan qayta qurilishi mumkin.

​
Recall: queries over the derived index

Recall should support:
  • lexical: “find exact terms / names / commands” (FTS5)
  • entity: “tell me about X” (entity pages + entity-linked facts)
    1. temporal: “11-noyabr 27 atrofida nima bo‘ldi” / “o‘tgan haftadan beri”
    1. opinion: “Peter nimani afzal ko‘radi?” 3. (ishonch + dalillar bilan)
  1. Qaytarish formati agentlar uchun qulay bo‘lishi va manbalarni keltirishi kerak:
  • kind (world|experience|opinion|observation)
  • Agar mualliflar bu haqda o‘ylashini istamasangiz: reflect vazifasi bu punktlarni jurnalning qolgan qismidan chiqarib olishi mumkin, ammo aniq ## Retain bo‘limiga ega bo‘lish eng oson “sifat richagi”dir.
  • entities (["Peter","warelay"])
    1. content (bayon qilingan fakt)
    1. source (memory/2025-11-27.md#L12 va h.k.)

​
10. Refleksiya: barqaror sahifalarni yaratish + e’tiqodlarni yangilash

  1. Refleksiya — rejalashtirilgan ish (kunlik yoki heartbeat ultrathink) bo‘lib, u:
    1. so‘nggi faktlar asosida bank/entities/*.md ni yangilaydi (entity xulosalari)
    1. mustahkamlash/qarama-qarshilikka qarab bank/opinions.md dagi ishonchni yangilaydi
    1. ixtiyoriy ravishda memory.md ga tahrirlarni taklif qiladi (“core-ish” barqaror faktlar)
  1. Fikr evolyutsiyasi (oddiy, tushuntiriladigan):
    1. har bir fikrda quyidagilar bo‘ladi:
      1. bayonot
      1. ishonch c ∈ [0,1]
    • last_updated
    • timestamp (manba kuni yoki mavjud bo‘lsa, ajratib olingan vaqt oralig‘i)
    1. yangi faktlar kelganda:
      1. entity ustma-ustligi + o‘xshashlik bo‘yicha nomzod fikrlarni topish (avval FTS, keyin embeddinglar)
      1. ishonchni kichik deltalar bilan yangilash; katta sakrashlar kuchli qarama-qarshilik + takroriy dalillarni talab qiladi

​
24. CLI integratsiyasi: mustaqil vs chuqur integratsiya

  1. Tavsiya: OpenClaw ichida chuqur integratsiya, lekin ajratiladigan yadro kutubxonasini saqlash.

​
26. Nega OpenClaw ga integratsiya qilish kerak?

    1. OpenClaw allaqachon biladi:
      1. ish joyi yo‘li (agents.defaults.workspace)
      1. sessiya modeli + heartbeats
      1. loglash + nosozliklarni bartaraf etish andozalari
    1. Asboblarni agentning o‘zi chaqirishini xohlaysiz:
    • openclaw memory recall "…" --k 25 --since 30d
    • openclaw memory reflect --since 7d

​
34. Nega baribir kutubxonani ajratish kerak?

    1. xotira mantiğini gateway/runtime bo‘lmasdan testlanadigan saqlash
  • dalil havolalari (tasdiqlovchi + qarama-qarshi fakt identifikatorlari)
  1. Tuzilishi:

​
Xotira asboblari kichik CLI + kutubxona qatlamidan iborat bo‘lishi mo‘ljallangan, ammo bu hozircha faqat tadqiqotdir.

  1. “S-Collide” / SuCo: qachon ishlatish (tadqiqot)
  2. Agar “S-Collide” SuCo (Subspace Collision) ni anglatadigan bo‘lsa: bu kuchli recall/kechikish muvozanatini maqsad qilgan ANN qidiruv yondashuvi bo‘lib, subfazolardagi o‘rganilgan/tuzilmaviy to‘qnashuvlardan foydalanadi (maqola: arXiv 2411.14754, 2024).
  • recall sifati leksik qidiruv bilan sezilarli darajada cheklanib qolganda
    1. boshlamang SuCo bilan.
    1. SQLite FTS + (ixtiyoriy) oddiy embeddinglardan boshlang; darhol UX yutuqlarining aksariga erishasiz.
      1. SuCo/HNSW/ScaNN turidagi yechimlarni faqat shunda ko‘rib chiqing:
      1. korpus katta bo‘lganda (o‘nlab/yuz minglab bo‘laklar)
      1. embeddinglarni qo‘pol kuch bilan qidirish juda sekinlashganda
  1. recall sifati leksik qidiruv bilan sezilarli darajada cheklanib qolganda
    1. Offline-ga qulay muqobillar (murakkablik ortib borish tartibida):
    1. SQLite FTS5 + metadata filtrlari (nol ML)
    1. Embeddinglar + brute force (bo‘laklar soni kam bo‘lsa, hayratlanarli darajada uzoq ishlaydi)
  • SuCo (tadqiqot darajasida; agar ichiga joylashtira oladigan mustahkam implementatsiya bo‘lsa, jozibali)
Ochiq savol:
  • sizning qurilmalaringizda (noutbuk + desktop) “shaxsiy yordamchi xotirasi” uchun eng yaxshi oflayn embedding modeli qaysi?
    • agar sizda allaqachon Ollama bo‘lsa: lokal model bilan embedding qiling; aks holda toolchain ichiga kichik embedding modelini yuboring.

​
Eng kichik foydali pilot

Agar minimal, lekin baribir foydali versiyani xohlasangiz:
  • bank/ entity sahifalarini va kundalik loglarda ## Retain bo‘limini qo‘shing.
  • Iqtiboslar bilan eslab qolish uchun SQLite FTS’dan foydalaning (yo‘l + qator raqamlari).
  • Faqatgina recall sifati yoki masshtab talab qilgandagina embedding qo‘shing.

​
Manbalar

  • Letta / MemGPT tushunchalari: “asosiy xotira bloklari” + “arxiv xotira” + vositalar orqali o‘zini tahrirlovchi xotira.
  • Hindsight Texnik Hisoboti: “retain / recall / reflect”, to‘rtta tarmoqli xotira, narrativ faktlarni ajratib olish, fikr ishonchliligining evolyutsiyasi.
  • SuCo: arXiv 2411.14754 (2024): “Subspace Collision” taxminiy eng yaqin qo‘shnilarni topish (ANN) orqali retrieval.