Question 1

Quali modelli locali dovrei eseguire e come l'hardware influisce su SecureContext?

Accepted Answer

SecureContext si degrada in modo graduale in base al livello di hardware — ogni livello alimentato da LLM si chiude silenziosamente (la funzionalità non contribuisce più) invece di interrompere l'ingestione o la ricerca. Ciò che cambia con l'hardware è la quantità di stack di intelligenza attivo:
| Tier | Hardware | Modelli compatibili | Cosa ottieni |
|---|---|---|---|
| Minimum | Qualsiasi CPU, ~2 GB RAM liberi | nomic-embed-text (solo embeddings) | Ricerca BM25+vector ibrida, memoria di lavoro, catena di audit, gating delle competenze — il nucleo. I livelli LLM (estrazione eventi, estrazione entità, adozione contraddizioni) rimangono dormienti. |
| Mid | 8–16 GB GPU (o Apple Silicon 16 GB+) | + qwen2.5-coder:14b o phi4:14b (uno alla volta) | + Estrazione fact-event al ingest (ragionamento temporale), adozione contraddizioni LLM, estrazione entità, riepiloghi file semantici L0/L1. |
| Full | 24 GB+ GPU (es. RTX 4090/5090) | + phi4:14b e gpt-oss:20b residenti insieme | Tutto, in parallelo, a latenza interattiva — più un potente generatore locale per QA/benchmark. Questa è la configurazione su cui sono stati misurati i delta dei benchmark pubblicati. |
Indicazioni per la scelta del modello (tutti misurati, vedere bench/): Embeddings: nomic-embed-text — richiesto, piccolo, funziona ovunque. Estrazione eventi (ZCEVENTEXTRACTMODEL, default phi4:14b): il nostro confronto ha mostrato phi4:14b con 100% di recall degli eventi / 100% di precisione data, accoppiando gpt-oss:20b a 2× la velocità. Su GPU più piccole, qwen2.5-coder:14b segue da vicino (84,6% di recall). Modelli di codifica ≠ migliori: qwen2.5-coder:32b ha ottenuto il peggior risultato (69,2%) nonostante sia il più grande.

Question 2

Cosa serve per eseguirlo?

Accepted Answer

Node 20+ e (consigliato) Docker per lo stack PostgreSQL + Ollama in bundle. L'installer a comando singolo fa tutto in circa cinque minuti. Un fallback SQLite funziona senza alcuna infrastruttura.

Question 3

Can multiple Claude Code sessions work on the same project without conflicts?

Accepted Answer

Yes — parallel sessions atomically claim tasks from a work-stealing queue (zero double-claims at 50 agents × 100 tasks in testing), coordinate through typed broadcasts (ASSIGN/STATUS/MERGE/REJECT), and keep private per-agent memory namespaces plus a shared pool. Department-style hierarchies (heads + workers, N-tier escalation) are supported for larger agent teams.

Question 4

Does SecureContext send my code or data to the cloud?

Accepted Answer

No. Memory, embeddings (Ollama nomic-embed-text), search, summarization, and the audit chain all run locally. It works fully offline (search degrades gracefully to keyword-only if Ollama is down) and costs $0 when idle.

Question 5

Are Claude Code skills safe to install?

Accepted Answer

Filesystem skills bundle scripts that run with your permissions, and Claude Code's native loader does not scan them. SecureContext adds the missing gate: AST scan at admission, HMAC verification before every execution, automatic quarantine on failure or post-admission change, and a verifiable chained log of every admission decision.

Question 6

Can it manage a long, multi-week project?

Accepted Answer

Yes — that's what the delivery layer is for. Define a program with phases (zc_program), and the open phase's acceptance checklist travels with the project state. Closing a phase requires an evidence table and generates a checkpoint document automatically. If the orchestrating agent dies mid-phase, a fresh one reconstructs the full delivery state from two calls — verified live.

Question 7

How much does SecureContext reduce Claude Code token costs?

Accepted Answer

Measured across a 10-session project: ~87% less context-overhead (session restore is ~1.5–4k tokens instead of 20–50k of re-pasted files). On long-lived projects with hundreds of accumulated memories, the recall budget keeps each recall at ~4k tokens where an unbounded dump would be ~47k. File questions are answered from 15–50× compressed semantic summaries instead of raw reads.

Question 8

How do I give Claude Code persistent memory across sessions?

Accepted Answer

Install SecureContext (one command — see Quick start), then agents call zcremember to persist decisions and gotchas and zcrecall_context({focus: "current task"}) at session start to restore them — ranked by relevance, with natural-language time queries answered directly from memory. Memory survives restarts, compactions, and reboots because it lives in PostgreSQL, not in the context window.

Question 9

How it's different from competitors

Accepted Answer

Most "memory for Claude Code" projects do one of three things: Wrap the conversation in JSON. Saves the messages, replays them. No semantic compression, no audit trail, no security model. Sync to a cloud service. Better UX but now your agent's tool history (potentially including secrets it touched) lives on someone else's server. Drop in vector search. Adds RAG over your docs but doesn't help with the actual problem — orchestrating what the agent does, verifying what ran, surviving session compaction. SecureContext is none of those: | | SecureContext | Cloud memory services | Vector RAG plugins | Native Claude Code | |---|---|---|---|---| | Persistent across restarts | ✅ Yes, local Postgres | ✅ But on their server | ⚠️ Only docs, not state | ❌ Lossy auto-compact | | Delivery programs + evidence-gated checkpoints | ✅ zcprogram | ❌ | ❌ | ❌ | | HMAC audit trail | ✅ Per-agent subkey | ❌ | ❌ | ❌ | | Per-agent identity isolation | ✅ HKDF + RLS | ❌ | ❌ | ❌ | | Skill admission gate (HMAC + AST) | ✅ | ❌ | ❌ | ❌ | | Tamper-evident chain + session replay | ✅ | ❌ | ❌ | ❌ | | Work-stealing queue for parallel sessions | ✅ FOR UPDATE SKIP LOCKED | ❌ | ❌ | ❌ | | Runs offline | ✅ Local Postgres + Ollama | ❌ | ⚠️ Depends | ✅ | | Cost when idle | $0 | Monthly subscription | $0 | $0 | Versus the well-known memory products: Mem0 and Zep are excellent hosted memory APIs for LLM apps you build; Letta (MemGPT) is an agent framework with self-editing memory. SecureContext targets a different job — me

Question 10

What's new (v0.46.1)

Accepted Answer

Temporal retrieval round 2 + the delivery-tool kit. Compound temporal questions decompose into event clauses with rank fusion; temporal searches render a chronological Timeline block and stale results carry age warnings. zcprogram adds multi-phase delivery programs with evidence-gated, auto-checkpointing phase closure — live-verified with an orchestrator handoff test (kill the orchestrator mid-phase; a cold replacement resumes from two calls). The embed lane now has a watchdog with budgeted self-healing. Recent releases: v0.46.0 — team memory (per-user keys, shared workspaces, write attribution), session replay with cryptographic provenance, OTel Gen-AI spans, CI/headless memory CLI, chunked embeddings, compliance report export. v0.44.0 — temporal fact supersession, durable PG task graph, first public LongMemEval numbers. v0.43.0 — token-budgeted recall digest (fixes the 47k-token recall firehose on mature projects). v0.41–0.42** — memory-quality program: focused recall, consolidation, bi-temporal windows, multi-hop graph retrieval, TTL memories, multimodal ingestion. Full details in CHANGELOG.md; architecture notes in ARCHITECTURE.md.

Question 11

NVIDIA GPU? add: -f docker/docker-compose.nvidia.yml

Accepted Answer

curl http://localhost:3099/health # → {"status":"ok","version":"0.46.1",...} docker exec securecontext-ollama ollama pull nomic-embed-text docker/start.ps1 / docker/start.sh wrap this with GPU auto-detection and a prod` mode (nginx reverse proxy).

Question 12

What is SecureContext?

Accepted Answer

If you use Claude Code on real projects, you have hit these walls: 1. Every session starts cold. You re-paste the same files, re-explain the same decisions, re-discover the same gotchas — and pay tokens for the same orientation work over and over. 2. Long projects have no delivery state. When an orchestrating agent dies mid-feature (crash, closed window, exhausted context), its replacement starts blind. People compensate with hand-written CHECKPOINT files. 3. There is no record of what agents actually did. When something goes wrong, Claude Code has no tamper-evident history of what ran. 4. Skill scripts run with your permissions, unscanned. Anyone (or any agent) with write access to ~/.claude/skills/ can plant code that runs on the next invocation. 5. Parallel sessions step on each other. Two windows claim the same task; one overwrites the other's edits. SecureContext is an MCP plugin that closes all five. Install it once and your agents get: Memory that survives restarts — decisions, gotchas, and project state restored in one call; per-file semantic summaries answer "what does this file do" at 15–50× compression instead of raw reads; temporal search that knows when things happened and flags stale knowledge. A delivery layer for long-running projects — zcprogram tracks multi-phase programs; closing a phase requires acceptance evidence and auto-generates a searchable checkpoint; a fresh orchestrator resumes a dead one's work from two calls. A cryptographic audit trail — every

designer-coordination-rules (Grade A)Competenza data-ai sicura per Claude AI. Classifica A. REGOLE DI COORDINAMENTO — estratte da roles.json deepPrompt per designer

Panoramica

Funzionalità chiave

Prezzi

Casi d’uso

Progettazione di una nuova funzionalità

Proposizione di un cambiamento di design

Pro & contro

Pro

Contro

Recensioni

Domande e risposte

Fai una domanda

Alternative a Competenze