Every April 2026 AI-news roundup lists the same releases. Fazm ships the one thing no roundup has, a parallel observer session that turns the firehose into personal memory.

The April 2026 cycle is dense. On the model side: Claude Opus 4.7 shipped GA on April 2, Haiku 4.5.2 followed as a minor refresh, Claude Mythos preview dropped April 7, Gemma 4 shipped as four Apache-2.0 variants (E2B effective 2B, E4B effective 4B, 26B MoE, 31B Dense) with multimodal inputs and native tool use, GLM-5.1 shipped as a 754B MoE, Qwen 3.6-Plus shipped with a 1M-token context, DeepSeek V3.2 cleared AIME at 92.7 percent, Grok 4.1 shipped with tighter agent tooling. On the runtime side: vLLM v0.18 added native gRPC serving behind --grpc, vLLM v0.19 flipped the async scheduler on by default and added complete Gemma 4 support, Ollama shipped tighter tool-use routing, and LM Studio acquired Locally AI on April 8. On the paper side: new arxiv work on long-context retrieval (1M+ token windows), agent verification (step-level success prediction), and efficient inference kernels. Every top SERP roundup lists these. None answers what a reader should do with the firehose.

A mechanism for converting the firehose into durable personal context. Top SERP roundups (artificialintelligence-news.com, the-decoder.com, venturebeat.com/ai, simonwillison.net, huggingface.co/papers) enumerate releases and papers as news items. None has a product that filters the stream against what the reader actually cares about and saves the result across sessions. Fazm has exactly that: a third ACP session named observer, pre-warmed alongside the main and floating sessions, driven by a system prompt that explicitly names itself a parallel intelligence, with save_observer_card, MEMORY.md writes, and skill-file creation as its output surface. That is an uncopyable implementation detail, not an editorial summary.

Two files. The system prompt lives at Desktop/Sources/Chat/ChatPrompts.swift lines 556-597, as a multi-line Swift string literal named chatObserverSession. The first sentence at line 557 reads verbatim: You are the Chat Observer, a parallel intelligence running alongside {user_name}'s conversation with their AI agent. You watch conversation batches and build persistent memory. The session registration lives at Desktop/Sources/Providers/ChatProvider.swift line 1050 as .init(key: "observer", model: "claude-sonnet-4-6", systemPrompt: chatObserverSystemPrompt), inside an array passed to acpBridge.warmupSession alongside the main and floating sessions. The underlying fan-out happens at acp-bridge/src/index.ts line 1320 via await Promise.all(toWarm.map(async (cfg) => { ... })).

Two reasons, both visible in the code. First, different system prompts. The main session has the full desktop-chat prompt with conversation history, tool inventories, and domain context. The observer has a 42-line focused prompt that says its one job is to watch and save. Sharing a session would mix those roles. Second, memory namespace. Each session id segments the ACP SDK's memory space (MEMORY.md + topic files under ~/.claude/projects/). The observer writes to the same project memory but with its own running context, so it is not distracted by the main chat's live tool invocations. The parallel Promise.all at acp-bridge/src/index.ts line 1320 is what makes three sessions warm cheap enough to be the default.

Four explicit tools plus the SDK's built-in memory system. From ChatPrompts.swift lines 569-582: save_observer_card for surfacing saved memories as auto-accepted inline cards (types insight, pattern, skill_created, summary), query_browser_profile for reading the locally extracted browser profile (identity, emails, accounts, tools), execute_sql for read access to app data and update access to ai_user_profiles, capture_screenshot capped at 1 per minute, and an explicit Skills workflow that creates files at ~/.claude/skills/{skill-name}/SKILL.md when a repeated workflow is detected 3+ times. The primary memory store is the SDK's own MEMORY.md plus topic files, not SQL.

Through an auto-accepted card surface. After the observer finishes a batch, acp-bridge fires a setChatObserverPollHandler callback registered at ChatProvider.swift line 1005. That handler calls pollChatObserverCards at line 3200, which reads rows from the observer_activity SQLite table where status = 'pending', injects them into the main chat as ObserverCardBlocks, immediately marks them as status = 'acted' with userResponse = 'approve' (line 3261), and fires PostHog observer_card_shown and observer_card_action events with auto_accepted: true (line 3274). The only user-visible button on the card is Deny, at line 3246. Click Deny and a rollback path fires; do nothing and the memory is permanent.

It is a single paragraph at ChatPrompts.swift lines 588-596 under the heading Rules, Be Conservative. The exact rules: quality over quantity, do NOT save routine queries or things already handled or temporary debug context or session-only info, DO save personal preferences, recurring patterns, important relationships, life events, professional context, communication style. Always check MEMORY.md first and skip near-duplicates. One memory plus one card per observation. Conclusions not narration: write 'Prefers X' not 'I noticed X'. Skills only for repeated workflows 3+ times, not preferences or one-off tasks. Think deeply, connect dots across sessions.

The April 2026 news cycle has roughly 15 major releases across models, runtimes, and tools in a single month. A single human cannot read every changelog, every arxiv abstract, every release blog and carry that state across weeks. A parallel observer session can. Example flow: you mention Opus 4.7 in a Wednesday chat, the observer writes 'User is evaluating Opus 4.7 for research tasks' to MEMORY.md and drops a card. Friday you ask the main session about Claude Mythos preview. The main session's prompt already includes MEMORY.md context, so the answer opens with 'Based on your Wednesday evaluation of Opus 4.7, here is how Mythos differs'. That is a news-filter behavior no news site can produce, because the filter is about YOU, not about the news.

The observer is a chat observer; it watches conversation batches. But because Fazm itself uses accessibility APIs across any Mac app (the same MCP servers booted in buildMcpServers at acp-bridge/src/index.ts), whatever you do in Chrome, Safari, Slack, Notion, Mail, Preview, a PDF reader, an RSS reader, or a native notetaking app feeds back into the main chat as context, which then passes through the observer. The observer's system prompt explicitly mentions query_browser_profile at line 574, which is a locally extracted index of identity, emails, accounts, tools, contacts, addresses, and payments. That is the connective tissue: your news reading surfaces in the main chat, the observer turns it into memory.

The memory lives in the SDK's project memory directory under ~/.claude/projects/, plus the observer_activity SQLite table, which is part of the local Fazm AppDatabase on disk. Nothing is uploaded to Fazm servers by the observer pipeline. The save_observer_card tool writes to a local SQLite row; pollChatObserverCards reads from that row. The only external event is a PostHog analytics ping (observer_card_shown / observer_card_action), and the event properties at line 3270 include activity_id, card_type, and the card content string, which is anonymized aggregate telemetry rather than model context.

Yes. Three commands close the loop if you have read access to the Fazm source. rg -n 'parallel intelligence' Desktop/Sources/Chat/ChatPrompts.swift locates line 557, the verbatim anchor sentence. rg -n 'observer.*claude-sonnet-4-6' Desktop/Sources/Providers/ChatProvider.swift locates line 1050, the session registration. rg -n 'Promise.all' acp-bridge/src/index.ts locates line 1320, the parallel fan-out. Zero install, four lines, full trace from system prompt to runtime behavior.

A newsletter pushes the same article to every reader. A smart feed reader ranks items you have not read yet. Both are editorial pipelines from news source to reader. The observer is the opposite direction: it reads the reader's conversations and distills what matters into personal memory that then biases every future answer. It has access to the actual content you discussed, not a clickstream. It persists across sessions via MEMORY.md, so context never resets. And it is tied to a concrete Claude-shaped session with a written system prompt that a user can read, not an opaque recommendation model.