Event

Boston Generative AI Meetup — AI Agents, Data & the Power of Context

When

Monday, 2026-11-30, 18:00–21:00 ET

Where

High Street Place Food Hall, 100 High Street, Boston, MA 02110

Host

Judah Phillips (founder; also Founder/Board Chair, Boston AI Week)

Sponsors

Squark, Boston AI Week, SmartCurrent

Admission

RSVP required; no price listed

Link

Event page

Four months after The Harness for AI Agents, same room-ish, the frame moves from harness to context: what the agent knows, how fresh it is, who may see it, and who audits it. The MongoDB pitch is that the context layer is a database category sitting beside the system of record. The interesting question is not whether that works but what invariant it claims to hold — staleness bound, entitlement filter, replayable provenance — and whether anything tests it.

Previous editions: 2026-07-23 — The Harness for AI Agents. The 2026-09-28 Boston AI Week edition (Laugh Boston) has no note here; see Boston AI Week 2026.

Research to align against, by theme:

Context engineering

agent-context thread (entry point: memory, harness, skills, who pays for the window) · context surfaces · Postman agents/APIs 2026-08-27 (compression as "a forgetting policy sold as an inference feature").

MCP and permissions

agent permission guardrails ("the bottleneck is permissions"; MCP authorization gap; DB-proxy gap) · AGNTCon/MCPCon NA 2026.

Memory and state

agent memory systems (Letta/Zep/Mem0; CLAUDE.md as schema) · memory as institutional knowledge · AICamp 2026-08-27 (Neo4j three-layer memory) · LangGraph checkpoints.

Retrieval

history as retrieval context (retrieve vs hold it all) · pocket-es (BM25) · ML design patterns (only hybrid-search mention) · SIGIR 2026 (rerankers) · LLM Engineer's Handbook (only MongoDB-as-vector-store).

Provenance and tracing

AICamp 2026-07-23 ("what is the provenance of a retrieved memory, and who audits") · agent telemetry systems.

Evaluation

Q3 agentic development (mutation-testing the reward) · harness first · visibility is verification. Draft: _drafts/2026-10-13-eval-construction-checklist.org (twelve nulls an eval has to reject).

Gaps this event could fill: no MongoDB note; no retrieval-eval-in-production note (recall@k, latency, cost); "context layer" as a product category unexamined.

• MongoDB SA, Boston, since ~2024-04; previously Moviri (2020–24) [unverified: LinkedIn snippet]. Background is performance engineering and capacity planning — co-presented Performance Engineering Maturity Model at CMG IMPACT2021. Latency/cost questions are fair game.
• GitHub: community-migrate-script (Community→Atlas sizing), rxc-seach-fe. Nothing agent-shaped public; the healthcare-benefits demo appears new.
• Nearest published analogue: Insurance Context Layer reference architecture — one live document per business object, agent memory beside it, hybrid search + rerank, Change Streams for freshness, immutable audit collection. "Field-level governance" named, mechanism not given. Also: why agentic AI needs a different data architecture (white paper).
• [ ]

"A context layer complements, not replaces, the system of record"

falsified if the demo writes decisions back only to the context layer, or if freshness depends on an unbounded Change Stream lag.

"Immutable audit collection"

falsified if a role with ordinary admin rights can update/delete it (i.e. immutability is convention, not enforcement).

"Hybrid search beats vector-only for this workload"

falsified absent a labeled set and recall@k for both; rhetoric without a table.

"Agent memory retained across interactions improves outcomes"

falsified if there is no ablation (memory on/off) on a fixed task set.

Field-level governance

falsified if entitlements are applied after ranking (leak via scores/snippets) rather than as a pre-filter.

1. Staleness: when a benefits rule changes in the system of record, what bounds how long an agent acts on the old context — measured SLO or best-effort Change Streams?
2. Provenance: does each memory carry a pointer to source document / version? Can you replay a past decision against the context as it was?
3. Validity: are invariants (plan-year eligibility, coverage limits) enforced by schema validation at write time, or by the model?
4. MCP permissions: is retrieval filtered by the end user's entitlements before or after ranking? How do you show $rankFusion can't leak a document the user can't see?
5. Retrieval eval: what's the ground truth for the benefits demo? How was $rankFusion vs $scoreFusion chosen, and is it re-run after an embedding-model change?
6. Memory conflict: episodic memory contradicts current policy — which wins, and is that rule written down and tested?
7. Audit: what stops an admin rewriting the immutable audit collection?
8. Where is MongoDB not the fit — graph-heavy entitlement reasoning, strictly serializable money movement, something else?
9. (Panel) Baier: revenue-per-employee is the pitch — what's the denominator when agents are counted as headcount?

• [ ] Recheck listing for the three panelists (week of 2026-11-23)
• [ ] Skim GAI Insights Substack for Baier's agent-adoption position
• [ ] Read the Insurance Context Layer reference architecture end to end
• [ ] Post-event: fill Notes by topic; slides per the AICamp 08-27 convention
• [ ] Consider a MongoDB / context-layer research note to close the gap