Lesson 77 — Intermediate

Memory & State Offloading

Sab kuch window mein rakhna zaroori nahi — kuch cheezein bahar store karo aur demand par wapas laao. Offload se window chhoti rehti hai, task lamba chalta hai, aur crash ke baad bhi progress bacha rehta hai.

Offload kya hai

Do alag sawaal hain jo ek hi jagah solve hote hain: kya yaad rakhna hai (memory — user prefs, decisions) aur kya window se hatana hai (offload — bhiyaan ka junk). Store common, lifecycle alag.

Active vs stored context

Har turn se pehle decide karo kaunsa data active window mein baithne layak hai:

• Active: isi step ka plan, current tool call, abhi ka evidence — andar.
• Stored: puraane decisions, user preferences, completed steps, intermediate artifacts — bahar, lazy load.
• Recall path: stored se wapas laane ka tareeka defined ho — keyword search, ID lookup, ya summary index.
• Eviction: turn khatam, raw junk hata do — memory ko TTL/dedupe se controlled rakho.

Note-taking agent

def start_turn(task):
    notes = read_notes()          # progress, decisions, open items
    return plan_next(notes, task) # agla step notes se, blank se nahi

def finish_turn(state, result):
    append_note(
        f"- done: {state.step} | next: {state.next} | "
        f"facts: {result.key_ids}")   # durable, chhota, specific
# Notes = agle session ka context — fresh ho tab likho, pressure mein nahi

Plans & progress persist karo

• Plan file: checklist + statuses (pending/done) — task list kabhi window mein nahi ghulti.
• Steps log: kya kiya, kya nahi — resume ke liye exact state.
• Intermediate results: content nahi, artifact paths — chhota pointer rakho.
• Decisions + why: "X choose kiya, kyunki Y" — baad mein defend karne layak.

Remember vs free — do alag kaam

Memory = value kal bhi hai (user preferences, learned facts). Offload = aaj ka junk hataana. Dono same store mein rah sakte hain, par unka retention alag hona chahiye.

• Memory entries: curated, deduped, TTL/review cycle — store chhota aur clean rakho.
• Offloaded scratch: task khatam, delete/archive — woh baad mein kaam ki nahi.
• Transparency: user-visible memory (kya yaad hai) + delete option — DPDP trust ka hissa (L8.6).
• Cost check: store ka size bhi kharcha hai — weekly prune karo.

Exercise

Common mistakes

• Offload karke poora re-read: Bahar likh kar har turn poora store load kar diya toh point khatam — window wahi ki wahi. Lazy recall sirf zaroori slice laao.
• Memory bina eviction: Store unlimited badhta rahe toh sasta storage bhi mehnga padta hai, aur recall bhi dheema. Dedupe + TTL mandatory.
• Notes vague: "Progress accha hai" type notes se resume nahi hota. Done, next, key facts — teen fields har note mein.

store = {"prefs": ..., "decisions": ..., "progress": ...}   # bahar
window = load(store, need=turn)                           # sirf zaroori andar