Research Practitioner Preprint · Not Peer Reviewed
The Consilium Project · Memory Architecture

The Cold Start
Problem in Multi-Agent Memory Systems

Every query begins with amnesia. Ten agents. No memory of yesterday. No record of tensions resolved, calibrations made, or users known. This is not a bug — it was a design choice. This paper maps what that choice costs, what adding memory risks, and why the correct question is not "how do we add memory" but "which memories, at which layer, with which expiry."

Author
J. Erickson
System
10 Agents / 18mo
Status
Active Build
Published
March 2026

When The Consilium receives a query, ten agents wake up with identical amnesia. Each carries its system prompt — its identity, its domain, its epistemic fingerprint — and nothing else. No record of what was asked yesterday. No accumulated understanding of the user. No memory of prior tensions or syntheses. The system prompt is the only persistent thing. Everything else is cold.

§1 The Network in a Stateless State

Stateless architecture was a deliberate choice made early in the system's design: simplicity of deployment on Cloudflare Workers, clean isolation between queries, elimination of stale context errors, and the interpretability benefit of knowing that any given output was produced solely from the current input plus the fixed system prompt.

Eighteen months of operation later, the costs are visible. Users re-establish context that was built in prior sessions. Agents have no memory of tensions they resolved. The system that analyzed three prior acquisitions for the same user treats query four as if queries one through three never happened.

// Agent Network — Stateless Query State 10 agents active · 0 shared memory

§2 What Statelessness Provides

The standard framing of stateless architecture focuses on what it lacks. The more useful framing is what it actively provides — because those provisions are at risk when memory is introduced.

2.1 Clean Agent Identity

Each agent has an engineered epistemic identity: a set of priors, a domain of authority, a characteristic type of skepticism. In a stateless system, the identity is perfectly consistent across every query. The contrarian is always the contrarian. The economist always enters the room with the same priors. Introduce persistent memory and the identity becomes subject to drift — an agent that has "seen" 200 prior queries about a market will develop implicit priors not in its system prompt. Those priors may be accurate, which looks like learning, or artifacts of query history, which looks like bias.

2.2 Query Independence

In the current system, the answer to query N has no causal relationship to the answer to query N-1. Errors in a prior session cannot compound into subsequent sessions. A confident-wrong answer on Monday does not contaminate Tuesday. In a system with persistent memory, a confident incorrect claim from a prior session could seed the context for future sessions — the temporal equivalent of within-session propagation, but potentially much harder to detect.

Statelessness is not a failure of ambition. It is a guarantee. The question is which guarantees you can afford to trade away, and for what.

§3 The Costs — Precisely Counted

Cost A · Overhead
User context re-establishment
Users re-explain their situation on return. In real estate analysis: restating property details, prior context, decision history. Estimated 15–30% of query content in returning-user sessions. Pure overhead — displaces analytical content without adding new information.
Cost B · Signal Loss
Lost calibration signal
When an agent's analysis is validated or invalidated by real-world outcomes, the current system has no mechanism to incorporate that signal. An economist systematically wrong about a market has no record of being wrong. The system prompt is static. There is nowhere for learning to live.
Cost C · Resolution Loss
Tension resolution evaporates
When agents reach Round 2 confrontation and one position wins, that resolution is lost at session end. The next session starts the same argument from scratch. High-severity tensions recur across sessions on similar queries — inefficient and potentially masking a solvable structural disagreement.
Cost D · Floor
Personalization floor
Without memory, every user interaction is equally generic. The system cannot develop a model of a specific user's risk tolerance, deal history, or decision-making patterns. The most obvious cost — and the one with the highest contamination risk if addressed naively.

§4 The Contamination Problem

The intuitive solution to cold start is to inject prior context into the agent's context window at session start. The technical implementation is straightforward. The contamination problem emerges from three sources that naive implementations consistently underweight:

01
Stale Priors
Memory from prior sessions carries conclusions that may no longer be valid. A market assessment from 90 days ago is not the same as one from today. Without a timestamp and decay function, it will be treated as current information — quietly poisoning the analysis without any visible signal that something is wrong.
Implemented — v0.3
02
Error Compounding
A confident-wrong output in session N becomes a prior in session N+1. Without outcome validation, the system has no mechanism to distinguish validated conclusions from unvalidated ones. In adversarial testing, confident incorrect claims propagated through the synthesis layer within a single session. In a stateful system, that propagation becomes temporal — crossing session boundaries undetected.
Open — Validation Layer Not Built
03
Identity Drift
Memories describing an agent's prior outputs cause that agent to anchor on its own prior positions — directly undermining the epistemic independence the system prompt establishes. An economist that remembers being bullish will be anchored toward continued bullishness, not because fundamentals support it, but because its own prior position is now in its context window. The memory creates the consensus pressure the architecture was designed to resist.
Implemented by design — agent memory isolation
proptechusa-memory / contamination-vectors.ts TypeScript
// The three contamination vectors — live status as of March 2026

// Vector 1: Stale Priors
const stalePriorRisk = {
  trigger:    'memory_age > decay_threshold',
  mitigation: 'timestamp all entries, apply relevance decay fn',
  status:     'IMPLEMENTED — proptechusa-memory v0.3'
}

// Vector 2: Error Compounding
// Confident-wrong in session N becomes prior in session N+1
// No mechanism to distinguish validated vs unvalidated conclusions
const errorCompoundingRisk = {
  trigger:    'unvalidated_memory injected as prior',
  mitigation: 'confidence tagging + outcome validation layer',
  status:     'OPEN — validation layer not yet built'
}

// Vector 3: Identity Drift (most counterintuitive)
// Agent memories of own prior outputs create anchoring
// Undermines epistemic independence that architecture requires
// Solution: NO agent-specific memory. Global user context ONLY.
const identityDriftRisk = {
  trigger:    'agent receives memory of own prior outputs',
  mitigation: 'agent memory isolation — user context only, never agent context',
  status:     'IMPLEMENTED by architectural design'
}

// The rule that follows from all three:
// Agents know what the USER brought to the system.
// Agents do NOT know what they previously said about it.

§5 The Layered Architecture

The design of proptechusa-memory reflects constraints derived from the contamination analysis. The architecture is not "add memory to agents" — it is "add a memory layer between sessions and agents, with explicit rules about what crosses the boundary."

// Context Window Composition — Three Architecture States Figure 1
Figure 1. Context window composition across three architecture states. Stateless (current) carries only system prompt and current query. Naive stateful injects full history, introducing contamination risk (shown in red). Target layered architecture adds curated user context while preserving agent identity isolation — the core architectural constraint driving the design of proptechusa-memory.
// proptechusa-memory Design Decisions
DecisionEnablesPreventsStatus
Global memory only — no agent-specificUser context continuityIdentity drift via self-anchoringIMPL
Timestamp + relevance decay on all entriesRecency prioritizedStale prior contaminationIMPL v0.3
Confidence tags on memory entriesUncertainty propagationUnvalidated → treated as factIN PROGRESS
Semantic relevance retrievalRelevant context without noiseFull history injection bloatPLANNED
Outcome validation layerReal-world calibration signalError compounding across sessionsOPEN

§6 Open Problems

Open · 01
Outcome Validation
The memory layer currently has no mechanism to distinguish validated conclusions from unvalidated ones. The architecture for connecting real-world deal outcomes to prior session memories has not been built. In the real estate context this is tractable — outcomes are observable. The feedback loop is not yet wired.
Open · 02
Semantic Relevance Retrieval
Current implementation retrieves memory by recency. Target implementation retrieves by semantic relevance to the current query — so residential history isn't injected when the query is commercial. The semantic retrieval layer adds latency and cost to session initialization. The tradeoff has not been characterized at production scale.
Open · 03
Decay Calibration
The relevance decay function uses a fixed half-life. The correct decay rate is almost certainly query-type-dependent: a market condition assessment has a shorter relevant lifespan than a user preference signal. Domain-specific decay calibration requires more operational data than is currently available.

§7 Conclusion

The cold start problem and the contamination problem are the same problem viewed from opposite sides. Every mechanism that reduces cold start introduces contamination risk. Every mechanism that prevents contamination reintroduces cold start. The design question is not which problem to solve but which tradeoff to accept at which layer.

The architecture described here accepts cold start at the agent identity layer — agents do not remember their prior outputs — in exchange for contamination protection. It accepts partial warm start at the user context layer in exchange for the re-establishment overhead cost. It defers the calibration and tension resolution problems to a future outcome validation layer that has not yet been built.

This is not a complete solution. It is an honest accounting of where a live system currently sits in a tradeoff space with no clean resolution.

The question was never
"how do we add memory."
The question is always "which memories,
at which layer, with which expiry."
Notes
[1]Within-session confident claim propagation documented in: Erickson (2026). "We Tried to Break Our Own AI." PropTechUSA.ai Research.
[2]Service binding architecture: Erickson (2026). "The Orchestrator Architecture."
[3]This is a practitioner preprint, not a peer-reviewed publication. The author is a self-taught developer with no formal research credentials. Findings are based on 18 months of production operation and offered as honest engineering documentation. Writing the paper while building the thing it describes.
J
Justin Erickson — Founder & CEO, PropTechUSA.ai
GED · Self-taught · Writing the paper while building the thing it describes · March 2026
Related research
// The system this paper describes
The Consilium

Ten agents. Stateless by design. Memory layer in active development. Ask it something hard.

Open The Consilium