Isifundo Sokukopisha Isinyathelo Ngesinyathelo Sokuqalisa I-GBrain: Isendlalelo Sezinkumbulo Ezizingenelayo Sakhiwe ngu-Y Combinator's Garry Tan wama-AI Agents.

Umenzeli wakho we-AI uhlakaniphile kodwa uyakhohlwa. Njalo iseshini entsha iqala kuqanda – akukho ukukhumbula ukuthi uhlangane nobani, ufunde ini, ukuthi unqume ini ngoLwesibili olwedlule. I-GBrain iwumthombo ovulekile wokulungisa lokho. Yakhelwe u-Garry Tan (uMongameli kanye ne-CEO ye-Y Combinator) ukuze inikeze amandla ukuthunyelwa kwe-OpenClaw kanye ne-Hermes, ungqimba lolwazi olweqa phansi kuqala, olusekelwa i-Postgres olungenisa imihlangano, ama-imeyili, ama-tweets, namanothi, bese ifaka ngokuzenzakalela igrafu yolwazi ethayiphiwe phezulu – ngaphandle kwezingcingo ze-LLM zokukhipha igrafu. Ubuchopho bokukhiqiza ngemuva kwama-ejenti angempela ka-Garry bukhona njengamanje Amakhasi ayi-146,646, abantu abangama-24,585, izinkampani eziyi-5,339, kanye ne-66 autonomous cron jobs. Kubhentshimakhi yayo (BrainBench, amakhasi angama-240 acebile-prose corpus), amahithi e-GBrain P@5 49.1% kanye no-R@5 97.9%ukuhola okungu-+31.4-point P@5 ngaphezulu kwe-codebase efanayo nesendlalelo segrafu sikhutshaziwe.

Lesi isifundo esisebenza ngezandla. Uzofaka i-GBrain endaweni, ungenise ifolda yamanothi amancane, wenze ukusesha kwangempela, ubukele intambo yegrafu yolwazi ngokwayo, bese uyixhuma ku-Claude Code nge-MCP. Cishe imizuzu engu-20 iqala ukuqeda. Yonke imiphumela yetheminali engezansi ithathwe ekufakweni bukhoma kwe I-GBrain v0.38.2.0. Indawo yokugcina (i-MIT-ilayisensi) ihlala ku-github.com/garrytan/gbrain.

Okwakhayo

Ekupheleni kwesifundo, uzoba:

• Indawo ~/.gbrain/brain.pglite database – ishumekiwe I-Postgres 17 (nge-WASM) nge-pgvector, i-zero server config.
• “I-brain repo” encane yamanothi e-markdown mayelana nabantu, izinkampani, nemiqondo.
• I-CLI yokusesha eyingxube esebenzayo ehlanganisa igama elingukhiye le-vector + BM25 + Reciprocal Rank Fusion (RRF), ene-ZeroEntropy reranker phezulu ngokuzenzakalelayo.
• Igrafu yolwazi ethayiphiwe (works_at, founded, invested_in, attended, advises, mentions) kukhishwe ngokuzenzakalelayo kumanothi akho.
• Iseva ye-MCP iyadalula 74 amathuluzi ngakho-ke i-Claude Code, i-Cursor, ne-Windsurf ingakwazi ukufunda nokubhalela ubuchopho ngokuqondile.

Okudingekayo

• macOS noma i-Linux (abasebenzisi beWindows: sebenzisa i-WSL2).
• Umhleli wekhodi.
• I-Bun ≥ 1.3.10 (isikhathi sokusebenza i-GBrain ivuliwe; i-repo's package.json imemezela lokhu njengenjini encane). Sizoyifaka kusinyathelo soku-1.
• Ukhiye wokushumeka we-API kusuka eyodwa kwe: ZeroEntropy (okuzenzakalelayo), OpenAI, noma Voyage. Ngaphandle kweyodwa, usengakwazi ukufaka nokusebenzisa ukusesha kwegama elingukhiye, kodwa gbrain query (i-hybrid + vector) ngeke ibuyise imiphumela.
• Ongakukhetha: ukhiye we-Anthropic API wokwandisa imibuzo eminingi ngesikhathi sokusesha.

Isinyathelo 1 – Faka i-Bun ne-GBrain

I-GBrain ibhalwe ku-TypeScript futhi isebenza ku-Bun. Yifake kuqala:

curl -fsSL  | bash
exec $SHELL                 # reload shell so `bun` is on PATH
bun --version

Manje faka i-GBrain. Kusukela ku-v0.38, indlela yokufaka ye-canonical iwukufakwa kwe-Bun eyodwa yomhlaba:

bun install -g github:garrytan/gbrain
gbrain --version
# gbrain 0.38.2.0

Isinyathelo sesi-2 – Qalisa ingqondo yakho

gbrain init --pglite inikeza isizindalwazi sendawo se-PGLIte ku ~/.gbrain/. I-PGlite igcwele i-Postgres ehlanganiswe ku-WASM – ayikho iseva, ayikho i-Docker, ilungile cishe imizuzwana emibili.

Kulesi sifundo sizohlehlisa umhlinzeki wokushumeka ukuze ukwazi ukulandela ngaphandle kokhiye we-API ngokushesha – sizokuxhuma ngezintambo kusinyathelo sesi-6 uma siqhuba usesho oluxubile:

gbrain init --pglite --no-embedding

(Uma ungathanda ukulungisa ukushumeka manje, setha okukodwa koku OPENAI_API_KEY, ZEROENTROPY_API_KEYnoma VOYAGE_API_KEY endaweni yakho ngaphambi kokusebenza kahle gbrain init --pglite.)

Okukhiphayo kwangempela kuthathwe ekufakweni okusha (okuncishisiwe ukuze kube mfushane – kukhona ukufuduka okungu-81 kusuka ku-schema v1 → v85):

Setting up local brain with PGLite (no server needed)...
  Schema version 1 → 85 (81 migration(s) pending)
  [2] slugify_existing_pages...
  [2] ✓ slugify_existing_pages
  [3] unique_chunk_index...
  [3] ✓ unique_chunk_index
  ...
  Brain ready at /home/you/.gbrain/brain.pglite
  0 pages. Engine: PGLite (local Postgres).

Manje unobuchopho obungenalutho. Qinisekisa:

gbrain stats
# Pages:     0
# Chunks:    0
# Embedded:  0
# Links:     0
# Tags:      0
# Timeline:  0

Isinyathelo sesi-3 – Dala irepo yobuchopho encane

I-repo yobuchopho iwuhla lwemibhalo nje lwamafayela okumaka. Ifayela ngalinye lilandela elika-GBrain iqiniso elihlanganisiwe + umugqa wesikhathi iphethini: isigaba samanje esiqondwa kakhulu phezulu, umkhondo wobufakazi besinezelo kuphela ngezansi.

Okubalulekile: ama-wikilinks kufanele asebenzise indlela ephelele ye-slug (isb, [[people/alice-chen]]hhayi nje [[alice-chen]]) ukuze isikhi segrafu sizixazulule. Lena i-gotcha yangempela – ngihlole womabili amafomu; ifomu elifushane likhiqiza ngokuthula izixhumanisi eziyiziro.

mkdir -p ~/my-brain/people ~/my-brain/companies ~/my-brain/concepts
cd ~/my-brain

Dala ikhasi lomuntu:

cat > people/alice-chen.md <<'EOF'
---
type: person
title: Alice Chen
tags: [founder, ai-infra]
---

Founder and CEO of [[companies/acme-ai]]. Previously staff engineer at
Google Brain. Focus area: inference optimization for small language models.

---

- 2024-03-12: Met at AI Engineer Summit. Discussed sparse MoE routing.
- 2024-09-04: Announced $12M seed led by Sequoia.
- 2025-01-18: Shipped open-source inference router on GitHub.
EOF

Ikhasi lenkampani:

cat > companies/acme-ai.md <<'EOF'
---
type: company
title: Acme AI
tags: [startup, inference]
---

YC W24 inference-optimization startup. Founded by [[people/alice-chen]].
Building latency-aware routing for sub-7B models.

---

- 2024-09-04: $12M seed, led by Sequoia.
- 2025-01-18: Open-sourced their inference router.
EOF

Futhi ikhasi lomqondo:

cat > concepts/inference-optimization.md <<'EOF'
---
type: concept
title: Inference Optimization
tags: [ml-systems]
---

Techniques to reduce latency and cost when serving language models:
quantization, speculative decoding, KV-cache reuse, and request batching.
EOF

Isinyathelo sesi-4 – Ngenisa i-repo

gbrain import ayinamandla (i-hash yokuqukethwe ikhishiwe). Sizodlula --no-embed ngakho-ke lesi sinyathelo siyanquma kubafundi abangakabi nalo isethi yokhiye wokushumeka — ukushumeka kugcwaliswa futhi kusinyathelo sesi-6. Okukhiphayo kwangempela:

gbrain import ~/my-brain/ --no-embed

[gbrain phase] import.collect_files start dir=/home/you/my-brain/ strategy=markdown
[gbrain phase] import.collect_files done 2ms files=3
Found 3 markdown files
[import.files] 3/3 (100%) imported=3 skipped=0 errors=0

Import complete (0.3s):
  3 pages imported
  0 pages skipped (0 unchanged, 0 errors)
  3 chunks created

Qinisekisa:

gbrain list
# companies/acme-ai           company   2026-05-22  Acme AI
# concepts/inference-optimization  concept  2026-05-22  Inference Optimization
# people/alice-chen           person    2026-05-22  Alice Chen

Isinyathelo sesi-5 — Faka igrafu yolwazi ngocingo

Ukuze uthole ukungenisa okokuqala, sebenzisa isikhipha isixhumanisi ngokusobala ukuze ugcwalise igrafu kusuka kuma-wikilink akho. Lokhu ukusho okumsulwa kwe-regex + okufakiwe – zero LLM izingcingo.

gbrain extract links --source db

Okukhiphayo kwangempela:

[extract.links_db] 3/3 (100%) done
Links: created 2 from 3 pages (db source)
Done: 2 links, 0 timeline entries from 3 pages

Imiphetho ethayiphiwe emibili ithathwe kuma-wikilinks: alice-chen --works_at--> acme-ai (kusuka “kuMsunguli kanye noMphathi Omkhulu we…”) kanye acme-ai --founded--> alice-chen (kusuka kokuthi “Yasungulwa ngu…”). I-inference cascade imililo ngokulandelana: FOUNDED → INVESTED → ADVISES → WORKS_AT → MENTIONS. Ayikho imodeli ku-loop.

Hlola igrafu ngokuqondile:

gbrain graph-query people/alice-chen --depth 1
# [depth 0] people/alice-chen
#   --works_at-> companies/acme-ai (depth 1)

gbrain backlinks companies/acme-ai
# [
#   {
#     "from_slug": "people/alice-chen",
#     "to_slug": "companies/acme-ai",
#     "link_type": "works_at",
#     "context": "Founder and CEO of [[companies/acme-ai]]...",
#     "link_source": "markdown",
#     ...
#   }
# ]

Lona umehluko phakathi kokusesha kwe-vector nokubuyiswa okuhlelekile. “Ubani osebenza e-Acme AI?” manje isiwukunqamula konqenqema lwe-hop eyodwa, hhayi amaphuzu afanayo. Leso siteshi sesakhiwo yisona esishayela ukuphakama kwe-+31.4-point P@5 ngaphezu kokuhlukile okukhutshazwe igrafu ku-BrainBench.

Isinyathelo sesi-6 — Yenza usesho

I-GBrain ithumela izenzo ezimbili zokusesha. gbrain search igama elingukhiye kuphela (BM25 ku-Postgres tsvector) futhi isebenza ngaphandle kokushumeka:

gbrain search "inference"
# [0.3648] companies/acme-ai -- YC W24 inference-optimization startup...
# [0.3648] people/alice-chen -- Founder and CEO of [[companies/acme-ai]]...

gbrain query iyipayipi elixubile eligcwele: i-vector (i-HNSW ku-pgvector) + BM25 + Reciprocal Rank Fusion + ukunwetshwa kwemibuzo eminingi ngokuzikhethela (Anthropic Haiku) + i-ZeroEntropy reranker ozikhethela yona. Idinga ukushumeka, esikuhlehlise esinyathelweni sesi-2 – kufake ngezintambo manje:

# Set one of: ZEROENTROPY_API_KEY (default), OPENAI_API_KEY, or VOYAGE_API_KEY
export OPENAI_API_KEY=sk-...
gbrain config set embedding_model openai:text-embedding-3-large
gbrain embed --all          # one-time backfill against your embedding provider
gbrain query "who works on small-model inference?"

# Set one of: ZEROENTROPY_API_KEY (default), OPENAI_API_KEY, or VOYAGE_API_KEY
export OPENAI_API_KEY=sk-...
gbrain config set embedding_model openai:text-embedding-3-large
gbrain embed --all          # one-time backfill against your embedding provider
gbrain query "who works on small-model inference?"

Izindlela zokusesha ezintathu ziphuma ebhokisini — conservative, balanced, tokenmax – ukuhlanganisa izindleko/ikhwalithi izinkinobho zibe ukhiye owodwa wokumisa. Okuzenzakalelayo ngu balanced ne-ZeroEntropy reranker ivuliwe. Ifomula ye-RRF: score = sum(1 / (60 + rank)).

Isinyathelo sesi-7 — Xhuma ku-Claude Code nge-MCP

Ubuchopho buwusizo kakhulu lapho i-ejenti ye-AI ikwazi ukuyifunda futhi ibhale kuyo ngokuqondile. U-GBrain uyadalula 74 amathuluzi phezu kwe-Model Context Protocol nge-stdio. Ukusethwa kwe-canonical umyalo owodwa (hhayi ifayela le-JSON elihlelwe ngesandla):

claude mcp add gbrain -- gbrain serve

Qinisekisa ukufakwa:

claude mcp list
# gbrain  stdio  gbrain serve

Manje buza uClaude Code into efana nale “sesha ingqondo ukuze uthole ukwaziswa okwengeziwe” futhi izodlula phakathi search ithuluzi bese ubuyisela imiphumela yakho ekhonjiwe. Amagama wangempela wamathuluzi e-MCP ayi-plain snake_case: get_page, put_page, delete_page, list_pages, search, query, add_link, get_backlinks, add_tagnokunye okungu-65.

Ikhesa ne-Windsurf sebenzisa ukulungiselelwa kwe-MCP JSON okujwayelekile kuma-UI ezilungiselelo zawo. I-server specs iyafana:

{
  "mcpServers": {
    "gbrain": { "command": "gbrain", "args": ["serve"] }
  }
}

Claude Desktop isebenzisa claude_desktop_config.json okwe i-stdio yendawo Amaseva e-MCP ane-JSON spec. Isilawuli kude Amaseva e-HTTP MCP kufanele engezwe ngokusebenzisa Izilungiselelo → Ukuhlanganiswa nethokheni yenkampani. Bheka docs/mcp/CLAUDE_DESKTOP.md ku-repo ye-GUI walkthrough.

Uma ufuna ukufinyelela ukude kusuka kunoma yimuphi umshini, shintsha i-stdio ye-HTTP:

gbrain serve --http --port 8787
# Bearer auth, default-deny CORS, two-bucket rate limit, per-request audit log.
# Postgres-only by design (PGLite is local-only).

Isinyathelo sesi-8 — Vumela ubuchopho buzisebenzele

I-GBrain ithumela iluphu ye-autopilot. Kusukela ku-v0.36.4, umyalo owodwa uhlanganisa uhlelo lokulungisa olu-odwe ukuncika, uhambisa isinyathelo ngasinye njengomsebenzi we-Minion, uhlole kabusha isikolo sempilo yobuchopho phakathi kwezinyathelo, futhi uyenqaba ukuchitha isikhathi esingaphezu kwesilinganiso sakho sezindleko:

gbrain doctor --remediate --yes --target-score 90 --max-usd 5

Noma isebenzise njenge-daemon:

gbrain autopilot --install        # cron-driven, 5-minute tick

Ubuchopho obunempilo bulala imizuzu engama-60 phakathi kwemikhaza. Abangenampilo bathola umjikelezo wasebusuku ogcwele: ukuvumelanisa, ukukhipha, ukushumeka, hlanganisa, hlanganisa. Izigaba ezintathu (synthesize, patterns, consolidate) avikelekile ukuze umenzeli oxhumeke ku-MCP angakwazi ukushisa buthule amakhredithi e-API.

Ngomsebenzi ongemuva we-ad-hoc, i Abangane ulayini uthatha imisebenzi yamagobolondo kanye nemisebenzi ye-LLM subagent ihlangene:

gbrain jobs submit sync --params '{}' --follow
gbrain jobs stats
gbrain jobs work --queue default

I-PGLite caveat eyodwa: gbrain jobs supervisor (i-daemon yesisebenzi eqala kabusha ngokuzenzakalelayo) i Ama-Postgres kuphela. Ilokhi yefayela ekhethekile ye-PGLIte ivimba inqubo yesisebenzi ehlukene – i-CLI iyenqaba ngephutha elicacile uma config.engine === 'pglite'. Uma uku-PGlite, namathela ku-inthanethi --follow imisebenzi okokufundisa, noma qhuba gbrain migrate --to supabase ngaphambi kokusukuma isisebenzi esiphikelelayo.

Umthetho womzila: umsebenzi onqumayo (donsa ama-tweets, hlaziya i-JSON, bhala ikhasi) uya kokuthi Minions; umsebenzi wokwahlulela (hlola ibhokisi lokungenayo, hlola okubalulekile) uya kubasebenzeli abancane be-LLM.

Okusanda kwenzeka, kumdwebo owodwa

markdown files  ──>  PGLite + pgvector  <──>  43 skills
(your repo,           (hybrid retrieval +     (HOW to use the brain;
 source of truth)      typed graph)           RESOLVER.md routes intent)
       ▲                                              │
       └──────────────  agent reads/writes  ──────────┘

I-markdown repo isistimu yerekhodi. I-GBrain iyisendlalelo sokubuyiswa + segrafu phezu kwayo. I-ejenti ifunda futhi ibhale kukho kokubili, futhi abantu bangahlala bevula noma iyiphi .md ifayela futhi ulihlele ngokuqondile — gbrain sync ithatha ushintsho.

Uzoya kuphi ngokulandelayo

• Ukuthwebula komugqa owodwa (okusha ku-v0.38): gbrain capture "the thought I want to remember" ihlala ngqo phakathi inbox/YYYY-MM-DD-. Iyamukela futhi --file, --stdinkanye nokungeniswa kwe-webhook nge gbrain serve --http /ingest.
• Thuthela eSupabase lapho ingqondo yakho ikhula ngaphandle kwendawo (i-PGLite inhle kufika kumakhasi angu-50K): gbrain migrate --to supabase.
• Ngenisa idatha yangempela ngenye yezindlela zokupheka: izwi (i-Twilio + OpenAI Realtime), i-imeyili + ikhalenda, abahlinzeki bokushumeka abangu-16, isango lokuqinisekisa.
• Qalisa amabhentshimakhi kuzelamani repo gbrain-evals: BrainBench (yokwenziwa) kanye gbrain eval longmemeval (ibhentshimakhi yomphakathi ye-LongMemEval).
• Bhala awakho amakhono. Ikhono ifayela lokumaka elinonile elifaka ukugeleza komsebenzi – izicupha, amasheke, isango lekhwalithi. gbrain check-resolvable iqinisekisa isihlahla sekhono ukuze sifinyeleleke / MECE / DRY.

Ukubheja okujulile ngemuva kwe-GBrain yilokho ihhanisi elincanyana, amakhono amafutha ishaya amakhono amancane ngemuva kwe-agent enamafutha. Isikhathi sokusebenza sihlala sincane; ubuhlakani buhlala emafayeleni e-markdown i-ejenti ewafundayo ngesikhathi sesinqumo. Ukuzibophezela ngakunye okwenzayo ku-repo yobuchopho kuwumongo unomphela umenzeli wakho awuzuza ngokuzayo uma evuka. Uma uyisebenzisa isikhathi eside, iba ngobuhlakani.

Isichazi Esibonakalayo sikaMarktechpost

# Install Bun
curl -fsSL  | bash
exec $SHELL

# Install GBrain
bun install -g github:garrytan/gbrain

gbrain --version
gbrain 0.38.2.0

# Defer embedding setup until search step
gbrain init --pglite --no-embedding
Setting up local brain with PGLite...
  Schema version 1 → 85 (81 migration(s) pending)
  [2] slugify_existing_pages... ✓
  [3] unique_chunk_index...     ✓
  ...
Brain ready at ~/.gbrain/brain.pglite
0 pages. Engine: PGLite.

--- people/alice-chen.md ---
type: person
title: Alice Chen
tags: [founder, ai-infra]
———

Founder/CEO of [[companies/acme-ai]].
Previously staff engineer at Google Brain.

———

- 2024-09-04: $12M seed, Sequoia led
- 2025-01-18: Open-sourced router

gbrain import ~/my-brain/ --no-embed
Found 3 markdown files
[import.files] 3/3 (100%) imported=3 skipped=0 errors=0

Import complete (0.3s):
  3 pages imported
  3 chunks created

gbrain extract links --source db
Links: created 2 from 3 pages
Done: 2 links, 0 timeline entries

gbrain graph-query people/alice-chen --depth 1
[depth 0] people/alice-chen
  --works_at-> companies/acme-ai (depth 1)

gbrain search "inference"
[0.3648] companies/acme-ai — YC W24
        inference-optimization startup...
[0.3648] people/alice-chen — Founder/CEO of
        [[companies/acme-ai]]...

gbrain query "who works on small-model inference?"
# Hybrid: vector + BM25 + RRF + ZeroEntropy rerank

claude mcp add gbrain -- gbrain serve

# For Cursor / Windsurf, use the standard JSON config:
{
  "mcpServers": {
    "gbrain": {
      "command": "gbrain",
      "args": ["serve"]
    }
  }
}

Okuthathwayo Okubalulekile

• I-GBrain (v0.38.2.0) inika abenzeli be-AI ungqimba lwenkumbulo oluqhubekayo, olufika phansi – olwakhiwa u-Garry Tan ukuze anikeze amandla ukuthunyelwa kwakhe kwe-OpenClaw/Hermes ephethe amakhasi ayi-146,646 nabantu abangu-24,585.
• Ukufaka kusebenza endaweni phakathi kwamaminithi angu-~30 ku-PGLIte (I-Postgres 17 ihlanganiswe ku-WASM, iseva enguziro) nezikali ku-Supabase noma ama-Postgres azibambele wona uma kudingeka.
• Yonke i-wikilink icutshungulwa yi-regex inference cascade (FOUNDED → INVESTED → ADVISES → WORKS_AT) ebhala imiphetho yegrafu ethayiphiwe ngamakholi we-LLM aziro.
• Ukusesha okuxubile (ivektha + BM25 + RRF + ZeroEntropy reranker) kushaya P@5 49.1% / R@5 97.9% ku-BrainBench — ukuphakama okungu-+31.4-point P@5 ngaphezu kokuhlukile okukhutshazwe igrafu.
• Idalula amathuluzi angama-74 nge-MCP — ifake ngentambo ku-Claude Code ngeyodwa claude mcp add gbrain -- gbrain serve futhi umenzeli wakho angakwazi ukufunda/ukubhala ubuchopho ngqo.

Hlola I-GitHub I-Repo futhi Amakhodi Okusebenzisa. Futhi, zizwe ukhululekile ukusilandela Twitter futhi ungakhohlwa ukujoyina wethu 150k+ ML SubReddit futhi Bhalisela ku Iphephandaba lethu. Linda! ukutelegram? manje ungasijoyina kuthelegramu futhi.

Udinga ukusebenzisana nathi ekuthuthukiseni i-GitHub Repo yakho NOMA Ikhasi Lobuso Lokugona NOMA Ukukhishwa Komkhiqizo NOMA I-Webinar njll.? Xhumana nathi