Engram's $98M Bet: Moving Enterprise AI Beyond RAG to Continuous Learning

Enterprise AI's Structural Dilemma: Why Models Remain Strangers After 500 Days

Enterprise work experience includes not only static documents but also project decision histories, long-term business preferences, and collaboration habits. Typical enterprise AI implementations rely on a loop: receive task → search/RAG → stuff context window → reason → session ends → rebuild context next session. This Inference-time Reading pattern incurs three costs:

Repeated query cost and latency : high-frequency tasks re-read the same materials, consuming tokens and inference time.

Context interference : large background documents dilute relevant signals in the context window.

Experience not automatically reusable : without a controlled learning mechanism, corrections and task trajectories don't carry over.

Core judgment: The value of enterprise long-term memory lies in letting models accumulate reusable knowledge across sessions, changing how they search, judge, and call tools on the next task.

Beyond RAG: A New Scaling Axis — Study Compute

General foundation models improve via more public data, pre-training compute, and test-time compute (reasoning chains). Engram proposes another scaling axis: Scaling Compute on Your Context — offline study compute on private enterprise data absorbed into parametric memory, yielding per-team intelligence.

This corresponds to an economic metric — Amortized Understanding :

Dimension                    | Traditional Enterprise RAG          | Engram's Study Compute
-----------------------------|-----------------------------------|----------------------------------
Knowledge understanding timing | Ad-hoc reading at inference time  | Continuous proactive study in background (training time)
Repeated query cost          | High-frequency tasks re-read same materials | Aims to reduce repeated reading cost
Inference latency            | Long context increases retrieval and reasoning time | Aims to reduce carrying irrelevant materials
Optimization target          | Expand context window (Context Size) | <strong>Intelligence-per-token</strong> (intelligence per token)

In high-frequency, repeated-query scenarios on the same large workspace, offline study front-loads understanding cost for reuse during online inference.

Internalize vs. Externalize: Three-Layer Collaborative Memory Architecture

Engram places Internalize and Externalize in the same system: stable, reusable patterns go into parameters or adapters; dynamic, precise, auditable facts stay in notes, databases, and retrieval tools.

Why Pure Text Memory Dies "Death by a Thousand Cuts"

Record all details : notes bloat, retrieval and reading costs grow.

Over-summarize : subtle judgment bases, cross-project links, and team preferences may be lost in compression.

Engram's Solution: Three-Layer Memory Synergy

┌─────────────────────────────────────────────────────────────┐
│ 1. Parametric Memory (Weights)                               │
│   · Carries: org long-term preferences, implicit association patterns, search & reasoning strategies, business intuition │
├─────────────────────────────────────────────────────────────┤
│ 2. Text Memory (Notes)                                       │
│   · Carries: editable, auditable high-frequency structured overviews (e.g., case status/project schema) │
├─────────────────────────────────────────────────────────────┤
│ 3. Raw Retrieval (Search & Tool / Raw Docs)                  │
│   · Carries: real-time changing facts, precise numbers, ERP state, verbatim citations │
└─────────────────────────────────────────────────────────────┘

Externalize : real-time dynamic numbers, ERP orders, regulation clauses — stay in databases and search tools for audit, correction, and rollback.

Internalize : stable organizational styles, professional judgment preferences, high-frequency business common sense — trained into model weights via lightweight parameter adapters (LoRA/PEFT).

Parametric memory provides search direction, structured notes provide editable views, raw retrieval verifies dynamic facts. The three layers work together, not replacing RAG with weights.

Harvey Joint Experiment: How Models Handle 100M Tokens of a Synthetic Law Firm

In August 2026, Engram and Harvey published the C&H synthetic law firm experiment: a virtual firm Calderwood & Harkness (C&H) with 266 cases, 9,286 documents, emails and contracts spanning years, totaling ~100M tokens. All evaluation tasks run in a shared, continuously evolving synthetic firm file system. The experiment mainly observes parametric memory's effect on agent search behavior; closed-book fact recall is only one part.

Metric                          | Base Model (Unlearned) | Engram-Studied Model
--------------------------------|----------------------|-------------------
First-turn precise search       | 3.2%                 | 99.2%
First-turn open full Notes      | 94%                  | 0%
First-turn mention correct related case | 0%            | 24.5%
Interactions to hit first correct case | 5.94 turns      | 3.73 turns

Base models typically open index notes first, then search broadly. Trained models more often construct targeted queries directly, using filter words not present in prompts or prior tool results, reaching relevant cases faster.

Efficiency and Cost in Synthetic Evaluation

Engram Agent (based on open-source Qwen3.8-27B + study adapter): per-task cost $0.13 , end-to-end full pass rate 30% .

Claude Opus 4.8 : per-task cost $1.32 , end-to-end full pass rate 25% .

Under this synthetic environment and pricing assumptions, Engram Agent achieves slightly higher full pass rate at lower per-task cost. This is an early result on a self-built benchmark and cannot be directly extrapolated to real law firms or other enterprise scenarios.

Turning Fragmented Enterprise Data into Training Signals

Enterprise documents cannot directly serve as high-quality training data. Feeding raw Word, Wiki, etc. into continued pre-training or simple QA fine-tuning fails to guarantee reusable task capabilities and increases hallucination, overfitting, and temporal governance risks.

Engram's key judgment: Raw Context must first be transformed into training tasks and verification signals. Their public offline study pipeline:

Enterprise native complex space (Docs + Wiki + Code + historical communications)
↓
1. Automated Structured Exploration (Study Rollouts): model autonomously browses file topology, discovers latent organizational patterns and business schemas
↓
2. Self-Generated Long-Horizon Tasks (Self-generated Tasks): based on discovered business patterns, synthesize cross-department, cross-dossier complex reasoning tasks
↓
3. Policy Distillation & Reinforcement Learning (On-policy Distillation + RL): Agent attempts problem-solving in sandbox, result verifier retains verified trajectories
↓
4. Parametric Internalization (PEFT / Knowledge Modules): solidify reusable patterns into lightweight parameter adapters

Continuous learning also requires tighter feedback loops between Research and Product. Production searches, corrections, and rejections can become candidate signals for the next offline study round, but must pass permission, quality, and safety validation first.

Three Remaining Challenges for Continuous Learning

Catastrophic Forgetting : Continuous weight updates on enterprise dynamics may damage the base model's general capabilities. Evaluating and controlling this degradation requires long-term validation.

Knowledge Freshness and "Right to be Forgotten" (GDPR/Unlearning) : Database permissions and files can be deleted directly, but information encoded in model weights is harder to precisely retract. Sensitive information internalization therefore needs stricter permissions, isolation, auditing, and machine unlearning mechanisms.

Multi-Tenant Continuous Training Engineering Complexity : If different departments and project teams own dynamic parameter modules, infrastructure must manage adapter versions, canary distribution, cold-start loading, GPU scheduling, and anomaly rollback. This is a distributed training and model ops problem.

Three Enterprise AI Lanes: Harvey, Glean, and Engram

The following comparison is this article's analytical framework, not an industry standard taxonomy. Each company places its primary capability in professional delivery, enterprise context, and continuous learning respectively:

Dimension              | Harvey (Vertical Professional Neo-Lab) | Glean (Horizontal Org Neo-Lab) | Engram (Continuous Learning Neo-Lab)
-----------------------|--------------------------------------|--------------------------------|------------------------------------
Strategic Entry Point  | Vertical deep water (legal/contracts) | Enterprise-wide collaboration entry (unified search) | Long-term memory & cognitive沉淀 (Memory/Learning)
Core Focus             | <strong>Job-first</strong> (professional outcomes) | <strong>Context-first</strong> (org relationships & permissions) | <strong>Learning-first</strong> (knowledge internalization & reuse)
Implicit Experience Carrier | Expert-written rubrics & evals | Enterprise graph, permissions, production context | Offline-study-formed parameter adapters (Weights)
Proprietary Model      | <strong>Tenet</strong> (complex legal delivery) | <strong>Waldo</strong> (high-frequency search planning & dispatch) | <strong>Knowledge Modules</strong> (long-term org intuition)
Memory Definition      | Case Vault & explicit lawyer preferences | Relationship graph & Harness policy memory | <strong>Parametric weights + structured notes + raw search synergy</strong>

Harvey already collaborates with Engram on long-term legal knowledge memory; Glean continues strengthening enterprise context, permissions, and governance. The three lanes can combine, but public materials don't yet prove convergence into a single architecture.

A possible combination:

Frontier Intelligence (general reasoning foundation)
+ Domain Intelligence (vertical delivery models, e.g., Harvey Tenet)
+ Enterprise Context & Governance (org graph & permission hub, e.g., Glean)
+ Firm-specific Learned Memory (long-term internalized enterprise intuition, e.g., Engram)

Five Takeaways for Chinese Enterprise AI Founders & Architects

Distinguish what to internalize vs. externalize : Dynamic, sensitive, volatile data stays in external tools; high-frequency, stable, reusable patterns reassessed for internal learning.

Calculate your business's "Cognitive Amortization Ledger" : If agents frequently and repeatedly query the same large private corpus, compare offline study cost vs. online repeated token, latency, and maintenance costs.

Measure memory effect by the next action : Beyond record counts, observe whether next-round tool calls, search keywords, evidence hit rates, and interaction turns improve.

Build Context-to-Task synthesis capability early : Before fine-tuning, establish synthesis and verification from Context to Task, avoiding feeding unprocessed unstructured materials directly as training samples.

Incorporate Human Correction into controlled feedback loops : User corrections on parameters and suggestions can become training signals, but are not directly usable training data — must pass permission, quality, and safety validation.

Engram currently offers a lane initially validated in synthetic enterprise environments: moving part of the understanding cost from inference time to learning time, and letting parametric memory, structured notes, and retrieval collaborate. Whether it holds stable in real, multi-tenant, continuously changing enterprise environments still depends on update, forgetting, permission governance, and model ops capabilities.