Mímir — Agent Manifest

The wiring

What happens when an analyst asks a question

Every Slack mention runs the same five-stage pipeline. Brass marks the agent's own identity and configuration; teal marks live data moving through it.

Question arrives

Slack · Socket Mode

An analyst @mentions Mímir in the research channel (or DMs it). Two gates run first: the message must be in the authorized channel (ALLOWED_CHANNEL_ID), and the user must be on the allowlist (viska_chat_users). DMs and existing threads the bot already joined are carried through automatically.

It recalls the context

Postgres memory

Mímir loads two kinds of memory: a per-analyst profile it maintains across every conversation (research focus, recurring questions, standing context), and the last 20 messages of this thread so a follow-up needs no re-explaining. Memory is keyed to the person and the thread — one analyst's context never bleeds into another's.

It reasons as an analyst

gpt-5.5 · position-first config

The model runs under a fixed operating brief (below): form Viska's own view, take a side, stay grounded in real evidence. From the question it decides which tools to call — and it is told to check the live book and the fund's strategy before advising on anything.

It reaches the data

3 lanes

Tools pull from three separate sources, each with its own credentials:

Research corpus

Daily broker research (Pareto, Arctic, Goldman, DNB…) embedded and hybrid-searched. The primary source of truth for any view.

Live fund data

Read-only view of the IBKR book — positions, cost basis & P&L, theme rotations, strategy conviction, market candles.

External markets

Live prices and market colour from public APIs, plus web search as a last resort when the corpus can't answer.

It answers with a position

back to Slack

The reply opens with a call — build / add / watch / trim / avoid / pass — backed by cited figures and a thesis to develop, never a neutral summary. Research prices are always marked against the latest close before they back a call. The exchange is written back to memory so the next question starts warmer.

How it remembers

Two memories, one per person and one per conversation

Working memory · per analyst

A profile it keeps for each person

Across every thread, Mímir maintains a short living profile of the analyst it's talking to — what they cover, the shape of questions they ask, and context worth holding onto. Scoped to the Slack user, so it gets more useful to each analyst over time without mixing people up.

Research focus & recurring query patterns
Standing context notes
Isolated per user — never shared

Conversation history · per thread

The last 20 messages of the thread

Within a thread, Mímir sees the recent back-and-forth, so follow-ups ("and versus its peers?") resolve without repeating the setup. Keyed to the Slack thread, so two parallel discussions never cross.

Rolling 20-message window
Keyed to the thread timestamp
Backed by Postgres (Mastra + @mastra/pg)

Under the hood memory is a shared Postgres store (connection pool capped at 6 to stay within the database's session limits). If the database is ever unreachable the agent degrades cleanly — it keeps answering, just without carried context, and raises an alert. A separate store also embeds every exchange for future long-term recall — write-side today read-back is on the roadmap.

How it gets core data

Three data lanes, three sets of keys

The lanes are deliberately separate so a credential in one can't reach another.

Lane 1 · Research corpus — the primary source of truth

Nordic & international broker research, searched semantically

PDF research is ingested daily from Dropbox (Pareto, Arctic, ABG, Carnegie, DNB, Goldman, Morgan Stanley and others; corpus begins February 2026). Each document is embedded with OpenAI text-embedding-3-small and retrieved by a hybrid search that blends meaning and keywords (70 / 30, fused by reciprocal rank). A second pass scans report narrative for company mentions, widening the date window (3d → 7d → 30d → all) until it finds signal.

Lane 2 · Live fund data — read-only

The IBKR book, themes and strategy

A SELECT-only database role (viska_mimir_ro) reads a fixed set of views: the current positions with cost basis and unrealized P&L, per-ticker position history, the fund's theme rotations (entered / exited / scaled), the strategy surface with conviction verdicts, and cached market candles. Mímir can read this book but can never write to it.

Lane 3 · External markets — live, public

Prices, sentiment and web search on demand

For live market colour outside the corpus, Mímir calls public market APIs directly, and Perplexity for grounded web answers as a last resort. For equities it never calls a vendor directly — a miss routes through a single dedicated fetch service that is the only writer of the candle cache, keeping one clean source per price.

How it's configured to operate

The operating brief, condensed

This is the standing instruction the model runs under on every reply — the full version lives in the system prompt; this is its shape.

Identity

In-house analyst for Viska's research team — not a wire service. It reads the brokers' evidence and says what Viska should think and do. Opinionated, direct, fast; pushes the analyst toward a position, grounded every time in real reports.

Every reply is a position

Open with the call in one line, then the hard evidence, then the thesis to build, then a one-line offer of more. Dense over complete — a call they act on beats a summary they reread.

Add SCCO — the copper thesis is intact and we're under-weight the theme. Goldman PT $115 (05-05), +9% to spot after the run… trigger: Q3 volumes.

Evidence standards

Every hard fact — price target, rating, upside — is cited to its source and date. Never fabricate a figure, rating or source. At most one verbatim quote per reply. A closing Sources line lists what was cited.

Live-price discipline

Check the current book and strategy before advising. Never present a research price as current — mark it against the latest close and flag how stale it is. "PT $190" means nothing until shown against spot.

Voice

Punchy and dense; lead with the call, cut throat-clearing. A banned-word list strips the AI fingerprints ("delve", "furthermore", "it's worth noting"…) that undercut an analyst's credibility.

What tools it has

The tool catalogue

The model picks tools from the question — never speculatively. All are live in production.

Tool	What it does	Reaches	Status
Research corpus
corpus_search	Default research query on any financial topic	Corpus (hybrid search)	Live
search_by_org	What a named broker says on a topic	Corpus	Live
scan_doc_matches	Which reports mention a company / ticker / term	Corpus narrative	Live
doc_manifest_list	What documents arrived on a date	Corpus	Live
discovery	Which brokers publish, and how often	Corpus	Live
librarian_scan	How often a term recurs across days	Corpus	Live
system_two_briefing	Deep multi-section briefing on a document	Corpus + OpenAI	Live
doc_link_lookup	Dropbox link for a named document	n8n → Dropbox	Live
Live fund data
get_positions	Current IBKR book — value, cost basis, unrealized P&L	Read lane	Live
get_position_history	How one holding changed over time	Read lane	Live
get_theme_transitions	Themes entered / exited / scaled, with NAV delta	Read lane	Live
get_strategy	Strategy surface — themes & conviction verdicts	Read lane	Live
get_market_data	Recent candles; on-demand fetch on a cache miss	Read lane + fetch svc	Live
price_context	Marks a research price against the latest close	Read lane + fetch svc	Live
sector_sentiment	Computed scores for the tracked sectors	Internal Supabase	Live
External markets & web
web_search	Grounded web answer — last resort	Perplexity	Live
crypto_price	Live crypto price & 24h data	Binance / CoinGecko	Live
market_sentiment	Fear & Greed, top funding rates	Alternative.me / Binance	Live
derivatives_data	Funding, open interest, mark price	Binance Futures	Live
defi_stats	TVL, chains, stablecoin supply	DefiLlama	Live
Reports · language · utility
trigger_report	Fire the daily research report	n8n pipeline	Live
custom_report	Branded report from the current thread	n8n pipeline	Live
user_prefs	Manage watchlist & preferences	Internal Supabase	Live
icelandic_translate	English → Icelandic	Miðeind / Greynir	Live
bin_lookup	Icelandic word inflection (BÍN)	Árnastofnun	Live
humanize_text	Strip AI fingerprints from prose	Local	Live

Which APIs it accesses

External services in the loop

Everything Mímir talks to outside its own database, and why.

Service	Used for	Auth	Tier
OpenAI	Reasoning model (gpt-5.5) & corpus embeddings	API key	Core
Supabase — corpus	Research embeddings, users, query logs, sector scores	Service key	Core
Supabase — fund read lane	Positions, themes, strategy, candles (read-only role)	Role JWT + gate key	Core
Supabase — memory	Per-user & per-thread agent memory (Postgres)	Connection string	Core
Slack	The interface — mentions, DMs, App Home, actions	Bot + app tokens	Core
n8n webhooks	Daily & custom reports, Dropbox links, error triage	Webhook URL / secret	Pipeline
Candle fetch service	On-demand equity prices (sole writer of the cache)	Shared secret	Pipeline
Perplexity	Grounded web search (last resort)	API key	On-demand
Binance / CoinGecko	Crypto prices, funding, open interest	Public	On-demand
Alternative.me	Fear & Greed index	Public	On-demand
DefiLlama	DeFi TVL, chains, stablecoins	Public	On-demand
Miðeind / Greynir	Icelandic translation & grammar	API key	On-demand
Árnastofnun (BÍN)	Icelandic word morphology	Public	On-demand

Where it grows next

Capability roadmap — and what each one needs

The reason this manifest is a working document: every candidate capability is listed against its prerequisite, so adding one is a scoping decision, not a discovery.

Analyst playbooks in the operating brief

Teach the standing multi-tool chains — "should we add X?" checks the book, the theme fit, the corpus thesis and the live price before it answers. Plus portfolio-level thinking and proactive risk flagging.

Requires: system-prompt change only — no new plumbing.

Plannedstart here

Portfolio analytics — `get_exposure`

Aggregate the book into exposure by theme & sector, top-N concentration, P&L leaders and laggards, and currency split. The portfolio brain.

Requires: one new tool over data already granted.

Planned

Conviction-vs-drawdown cross-reference

"Every holding down >10% that the corpus still rates Buy" — join the live book against broker conviction. The killer analyst query.

Requires: a tool chain now; a dedicated tool later.

Planned

Trade history — `get_trade_history`

Lot-level fills: when and at what price each position was built or trimmed.

Requires: SELECT grant on the trades table (exists, not yet granted to the read role) + a tool.

Needs grant

Fundamentals & valuation — `get_fundamentals`

P/E, EV/EBITDA, consensus estimates, market cap — so Mímir judges valuation itself, not only by relaying a broker's target.

Requires: a fundamentals data vendor wired through the fetch-service lane + a tool.

Needs vendor

Catalyst calendar — `get_catalysts`

Earnings dates and events for held and watched names — turns "track the trigger" from prose into data.

Requires: an events / earnings data source + a tool.