Yakha Inkumbulo Eqinile, Ehleliwe Yama-Agent e-AI nge-loci

Yini i-loci?

loci iyisendlalelo senkumbulo yesikhathi eside esinezinhlobo sama-agent e-AI. Ikhipha amaqiniso ahlala isikhathi eside ezingxoxweni, iwagcine ku-SQLite, iwahlele ngokubaluleka ukuze akhunjulwe kamuva, futhi isuse ulwazi olungasasizi.

Le phrojekthi isekelwe embonweni wokuthi izinkumbulo zidinga kokubili indawo nesakhiwo. Esikhundleni sokuphatha zonke izitatimende ezigciniwe ngendlela efanayo, i-loci ihlukanisa izintandokazi, izikhathi zokugcina, izinhloso, ubudlelwano, nezinye izinhlobo zolwazi ukuze zilandele izikhathi zokuphila ezifanele.

I-loci yathathwa ohlelweni lwenkumbulo olusetshenziswa ekukhiqizeni yi-Lima, yaphinde yabhalwa ukuze izimele kunoma yiluphi uhlelo oluthile oluyisingathi.

Indlela uhlelo lwenkumbulo olusebenza ngayo

Izinkumbulo ezinezinhlobo

Yonke inkumbulo ingeyesinye sezigaba eziyisishiyagalombili:

• routine
• study
• preferences
• finance
• goals
• relationships
• constraints
• ephemeral

Izinkumbulo nazo zinamazinga amane okubaluleka kanye nomjikelezo wesimo. Lesi sakhiwo sivumela uhlelo ukuthi luphathise intandokazi ehlala isikhathi eside ngendlela ehlukile kunesikhathi sokugcina sesikhashana noma iqiniso elihlala isikhathi esifushane.

Izindlela ezimbili zokukhipha ulwazi

Ukukhipha ulwazi kuhlanganisa izindlela ezimbili. I-extractor esekelwe kumodeli ifunda ingxoxo ukuze iqonde incazelo, kuyilapho ama-heuristic e-regular expression ethola imisho esekelwayo eshiwo ngokusobala. Imiphumela iyahlanganiswa, kuthi okukhiqizwe imodeli kube yikona okubalulekile.

Indlela ye-regex ingaqhubeka nokusebenza lapho umhlinzeki wemodeli engatholakali. Izingcingo zamamodeli, ama-SDK abahlinzeki, namakhi e-API kushiyelwa ngamabomu uhlelo oluyisingathi, olungafaka i-ModelExtractor.

Ukuhlunga, ukuhlela ngokubaluleka, nokuncipha

Isihlungi sokubhala sivimbela imizwa yesikhashana nezitatimende ezifanayo ezihlala isikhathi esifushane ukuthi zibe idatha yephrofayela engunaphakade. Ngokwesibonelo, isitatimende esithi “Ngikhathele namuhla” singaba yiqiniso amahora ambalwa kodwa sidukise ngemva kwezinyanga.

Lapho izinkumbulo zengezwa kumongo wemodeli, i-loci ikhawulela umphumela kumaqiniso ayishumi nambili. Iwahlela ngokubaluleka, bese icabangela ukusondela kwesikhathi sokugcina kanye nokusha kolwazi. Lo mkhawulo uvimbela i-prompt ukuthi ikhule ingapheli.

Ukuhlanganiswa kwansuku zonke nokuncipha kugcina indawo yokugcina ilusizo:

• Amakhophi aphindaphindwe ngokuphelele ayahlanganiswa.
• Ukuqagela okungaqinisekisiwe kugcina kuphelelwa yisikhathi.
• Amaqiniso esikhashana anesikhathi sokuphila samahora angu-24.
• Izinhloso eziqediwe ziphelelwa yisikhathi ngemva kwezinsuku ezingu-90.

Khetha ukuhlanganiswa

I-loci inikeza umnyombo owodwa owabiwe nama-adapter amabili.

Iseva ye-MCP

I-adapter ye-MCP isebenza nanoma iyiphi iklayenti le-MCP elihambisanayo. Iveza amathuluzi amathathu: remember, recall, kanye ne-forget.

I-MCP isekelwe ezicelweni nasezimpendulweni, ngakho ayinayo i-hook ezenzakalelayo yokuphela kohambo. I-agent kufanele ibize i-remember ngenhloso. Inkumbulo ingabuyiswa futhi ifakwe kumongo uma iceliwe.

I-plugin ye-Hermes

I-adapter ye-Hermes inikeza amathuluzi amathathu afanayo, kodwa futhi isekela ukukhipha okuzenzakalelayo futhi ifake umongo ngaphambi kwazo zonke izingcingo. Lokhu kungenzeka ngoba i-Hermes inikeza i-hook ye-post_llm_call engahlola ingxoxo ngemva kokuba imodeli iphendulile.

Ukuhlanganiswa okuqondile kwe-Python

Sebenzisa umtapo wolwazi oyisisekelo ngokuqondile lapho ulawula umjikelezo we-agent futhi ufuna ukunquma ukuthi ukukhipha, ukugcina, ukubuyisa, nokufaka umongo kwenzeka nini.

Ukufaka nokusetha

Faka iphakheji eyisisekelo

pip install mowave-loci

Faka ukwesekwa kwe-MCP

pip install mowave-loci[mcp]

Ngemva kokufaka i-MCP extra, engeza i-loci ekucushweni kweklayenti lakho le-MCP:

{
  "mcpServers": {
    "loci": {
      "command": "python",
      "args": ["-m", "loci.adapters.mcp_server.server"]
    }
  }
}

Qala kabusha noma ulayishe kabusha iklayenti ukuze lithole amathuluzi e-remember, recall, kanye ne-forget.

Faka i-plugin ye-Hermes

hermes plugins install <you>/loci
hermes plugins enable loci

Faka esikhundleni se-<you>/loci nge-repository ye-plugin efanele noma umthombo wokufaka kwakho.

Lungiselela indawo yokugcina

Ngokuzenzakalela, i-loci igcina izinkumbulo kusizindalwazi se-SQLite ku-~/.loci/memory.db. Setha i-environment variable ethi LOCI_DB ukuze usebenzise elinye ifayela.

export LOCI_DB=/path/to/memory.db

Akudingeki iseva yesizindalwazi ehlukile.

Ukusetshenziswa okuyisisekelo ku-Python

Ukugeleza komsebenzi okuyisisekelo ukudala indawo yokugcina, ukukhipha izinkumbulo ezingaba khona, ukuzihlunga, ukugcina ezamukelwe, bese wengeza izinkumbulo ezisebenzayo ku-system prompt.

from loci import Store, extract, should_persist, context_block

store = Store()

for memory in extract(
    "eu prefiro respostas curtas",
    assistant_text=reply,
):
    if should_persist(memory):
        store.upsert(memory)

system_prompt += "\n\n" + context_block(store.active())

Kulesi sibonelo, i-reply imele umbhalo womsizi ohambisana nengxoxo. Umlayezo womsebenzisi wesiPutukezi usho ukuthi umsebenzisi uthanda izimpendulo ezimfushane.

1. Store() ivula indawo yokugcina izinkumbulo ye-SQLite elungiselelwe.
2. extract(...) ikhiqiza izinkumbulo ezingaba khona engxoxweni.
3. should_persist(memory) isebenzisa isihlungi sokubhala.
4. store.upsert(memory) ifaka noma ibuyekeze inkumbulo eyamukelwe.
5. store.active() ithola izinkumbulo ezisebenzayo.
6. context_block(...) ifometha lezo zinkumbulo ukuze zifakwe ku-prompt.

Ukusebenzisa amathuluzi e-MCP

Uma iseva ye-MCP isilungisiwe, i-agent exhunyiwe ingasebenzisa imisebenzi emithathu:

• remember: gcina ngamabomu iqiniso okufanele ligcinwe.
• recall: thola ulwazi olugciniwe lapho lufaneleka khona.
• forget: susa ulwazi okungasafanele lugcinwe.

Ngenxa yokuthi i-adaptha ye-MCP ayizikhiphi ngokuzenzakalelayo izinkumbulo, imiyalelo ye-ejenti kufanele icacise ukuthi i-ejenti kulindeleke nini ukuthi ibize i-remember. Ungacabangi ukuthi wonke umjikelezo wengxoxo oqediwe uzohlolwa bese ugcinwa.

Amathiphu athuthukile okuhlanganisa

Gcina ingqondo yomhlinzeki ingaphandle komongo oyinhloko

Umongo oyinhloko we-loci usebenzisa kuphela umtapo ojwayelekile we-Python. Uhlelo lwakho luphethe izingcingo zemodeli, okhiye be-API, nama-SDK abahlinzeki. Uma ukukhishwa kwencazelo kudinga imodeli, faka i-ModelExtractor evela ku-host esikhundleni sokwengeza ukuncika kwabahlinzeki kumongo oyinhloko.

Qalisa imisebenzi yomjikelezo wokuphila ku-host

i-loci ichaza ukuma kwememori, ukuhlanganiswa, nokuziphatha kokuncipha, kodwa ayihleleli umsebenzi wansuku zonke. Sebenzisa i-cron, i-s6, i-systemd, noma isihleli esilinganayo se-host ukuze uqalise imisebenzi yansuku zonke.

Gcina umkhawulo womongo

Gwema ukweqa ibhulokhi yomongo ehlelwe ngokwezinga ngokwengeza yonke imininingwane egciniwe kuwo wonke ama-prompt. Umkhawulo wamaqiniso ayishumi nambili ukhona ukulawula ukukhula komongo, kuyilapho ubeka phambili ulwazi olubalulekile, olufike ngesikhathi, nolwakamuva.

Sebenzisa i-adaptha efanele

• Khetha i-MCP lapho ukuphatheka phakathi kwama-client e-MCP kubalulekile futhi imisebenzi yememori ecacile yamukelekile.
• Khetha i-Hermes lapho ufuna ukukhishwa okuzenzakalelayo ngemva kwezingcingo zemodeli nokufakwa kwememori ngaphambi kwengcingo ngayinye.
• Khetha i-direct Python API lapho uhlelo lwakho luphethe umjikelezo we-ejenti futhi ludinga ukulawula okunembile kokuphikelela nokwakhiwa kwe-prompt.

Vumela i-host iphathe ezokuthutha namaseshini

Umjikelezo we-ejenti, ukusatshalaliswa kwamathuluzi, amaseshini, ezokuthutha, nokuqinisekisa kungaphandle kwesibopho somongo oyinhloko. Ama-host afana no-Hermes, OpenClaw, noma u-Claude Code aphatha isikhathi sokusebenza esizungezile, kuyilapho ama-adaptha ephatha ezokuthutha nokuqinisekisa okuqondene ne-host.

Qalisa isethi yokuhlola

pytest -q

Le phrojekthi inezivivinyo ezingu-92 ezisebenza phezu kwesitoreji sangempela se-SQLite efayelini lesikhashana. Isitoreji uqobo asilingiswa, okusiza ukuqinisekisa ukuziphatha kusetshenziswa ubuchwepheshe obufanayo bokugcina obusetshenziswa ekusetshenzisweni kwangempela.

Isiphetho

i-loci inikeza ama-ejenti e-AI ungqimba lwememori oluncane, olunezinhlobo, noluhlala isikhathi eside ngaphandle kokudinga iseva ehlukile. Umongo wayo owabiwe uphatha ukwakheka kwememori nomjikelezo wokuphila, kuyilapho i-MCP, i-Hermes, noma i-ejenti yakho ye-Python inikeza isikhathi sokusebenza esizungezile. Qala nge-adaptha ehambisana ne-host yakho, gcina izibopho zokukhipha nokuhlela zicacile, futhi usebenzise ukufakwa komongo okuhlelwe ngokwezinga ukuze uvimbele imemori ekugcwaleni i-prompt.