Agentic Memory for Scientific Discovery

A Tutorial on Using Persistent Memory in AI-Powered Research Workflows

---

🎯 What is Agentic Memory?

Agentic memory enables AI agents to remember, learn, and build knowledge across sessions—transforming them from stateless tools into cumulative research collaborators.

The Problem

Traditional AI agents are like researchers with amnesia: - ❌ Forget everything between sessions - ❌ Can't learn from past experiments - ❌ Can't connect findings across time - ❌ Can't provide historical context

Result: You spend time explaining context every session.

---

The Solution

With agentic memory, agents become knowledge-building partners: - ✅ Remember experimental results - ✅ Learn what works (and what doesn't) - ✅ Connect patterns across studies - ✅ Provide full provenance for every claim

Result: Agents that get smarter over time, just like human researchers.

---

🧬 Why This Matters for Genomics

Use Case 1: Novel Isoform Discovery

Without Memory:

Day 1: "I found a novel splice site in BRCA1 exon 11"
Day 2: Agent forgets → starts validation from scratch
Day 3: Repeat validation → waste time

With Memory:

Day 1: Agent stores discovery + evidence
Day 2: Agent remembers → "I validated this last week, confidence 0.92"
Day 3: Agent builds on it → "Let's check nearby exons for similar patterns"

---

Use Case 2: Experiment Tracking

Without Memory:

Researcher: "Did adding chromatin modality help?"
Agent: "I don't have that information"
Researcher: *Searches through 50 log files manually*

With Memory:

Researcher: "Did adding chromatin modality help?"
Agent: "Yes. In 3 experiments (Feb 2026), F1 improved by 8% (0.87→0.91).
        Especially effective for tissue-specific splicing.
        Evidence: experiments/chromatin_ablation.json"

---

Use Case 3: Literature Knowledge

Without Memory:

Researcher: "Has anyone found cryptic sites in BRCA1 exon 11?"
Agent: *Searches PubMed again (5 minutes)*

With Memory:

Researcher: "Has anyone found cryptic sites in BRCA1 exon 11?"
Agent: "Yes. Wang et al. (2023, Cell) reported 127 novel junctions 
        including BRCA1 exon 11. I stored this in 
        memory/literature/Wang2023Cell.md last month.
        Would you like the full citation?"

---

🏗️ How It Works

Architecture Overview

┌─────────────────────────────────────────────┐
│         Your Research Workflow              │
│  (Run experiments, analyze data, write)     │
└──────────────────┬──────────────────────────┘
                   │
                   │ stores results
                   ▼
┌─────────────────────────────────────────────┐
│           Agentic Memory Layer              │
│                                             │
│  ┌──────────────┐  ┌──────────────┐       │
│  │  Experiments │  │  Discoveries │       │
│  │     (.md)    │  │     (.md)    │       │
│  └──────────────┘  └──────────────┘       │
│                                             │
│  ┌──────────────┐  ┌──────────────┐       │
│  │  Literature  │  │  Hypotheses  │       │
│  │     (.md)    │  │     (.md)    │       │
│  └──────────────┘  └──────────────┘       │
└──────────────────┬──────────────────────────┘
                   │
                   │ traces to
                   ▼
┌─────────────────────────────────────────────┐
│          Raw Data & Artifacts               │
│  (logs, configs, papers, plots, tables)     │
└─────────────────────────────────────────────┘

Key Principle: Memory preserves the full chain from claim → evidence → raw data.

---

Memory as Markdown Files

Memory is stored as human-readable Markdown files (not opaque databases):

memory/
├── experiments/
│   ├── chromatin_modality_2026Q1.md
│   └── histone_marks_ablation.md
├── discoveries/
│   ├── brca1_novel_isoforms.md
│   └── tp53_cryptic_sites.md
├── literature/
│   ├── wang2023_cell_splicing.md
│   └── key_papers_alternative_splicing.md
└── hypotheses/
    └── high_delta_predicts_novel.md

Benefits: - ✅ Human-readable: You can open and read any memory file - ✅ Git-versionable: Track how knowledge evolves over time - ✅ Editable: Correct mistakes or add notes - ✅ Transparent: See exactly what the agent remembers

---

📝 Example: Storing an Experiment

Step 1: Run Your Experiment

from agentic_spliceai import MetaModel

# Run experiment
model = MetaModel(modalities=["base", "sequence", "chromatin"])
results = model.train(data="chr21", epochs=50)

print(f"F1 Score: {results.f1}")  # 0.91

---

Step 2: Store in Memory

from agentic_spliceai.memory import AgenticMemory

memory = AgenticMemory()

# Store experiment
memory.store_experiment({
    "date": "2026-02-15",
    "hypothesis": "Chromatin modality improves tissue-specific predictions",
    "config": {
        "model": "meta-v2",
        "modalities": ["base", "sequence", "chromatin"],
        "test_set": "chr21"
    },
    "results": {
        "f1": 0.91,
        "precision": 0.89,
        "recall": 0.93
    },
    "conclusion": "Chromatin effective (+8% F1 improvement)",
    "artifacts": [
        "logs/chromatin_exp.log",
        "plots/chromatin_roc.png"
    ]
})

---

Step 3: Memory File Created

File: memory/experiments/chromatin_modality_2026Q1.md

# Chromatin Modality Experiment (Q1 2026)

## Experiment: chromatin-chr21-validation
**Date**: 2026-02-15  
**Hypothesis**: Chromatin modality improves tissue-specific predictions

### Setup
- Model: meta-v2
- Modalities: base + sequence + chromatin (ATAC-seq)
- Test set: chr21 validation

### Results
- **F1**: 0.91 (baseline: 0.83, improvement: +8%)
- **Precision**: 0.89
- **Recall**: 0.93

### Conclusion
Chromatin accessibility significantly improves predictions for 
tissue-specific splice sites. Most effective in genes with 
dynamic chromatin state (e.g., immune response genes).

### Evidence
- [logs/chromatin_exp.log]
- [plots/chromatin_roc.png]
- [configs/chromatin_exp.yaml]

### Related
- See: memory/modality_effectiveness/chromatin.md

---

🔍 Example: Querying Memory

Natural Language Queries

# Ask a question
results = memory.query(
    "What have we learned about chromatin modality effectiveness?"
)

# Agent finds relevant memories
for result in results:
    print(f"Category: {result.category}")
    print(f"Finding: {result.content}")
    print(f"Evidence: {result.links}")
    print("---")

Output:

Category: experiments
Finding: Chromatin modality improved F1 by 8% (0.83 → 0.91) in chr21 validation
Evidence: [logs/chromatin_exp.log, plots/chromatin_roc.png]
---

Category: modality_effectiveness
Finding: Chromatin most effective for tissue-specific splicing (F1=0.89)
         Least effective for housekeeping genes (F1=0.71)
Evidence: [47 experiments across 12 tissues]
---

Category: literature
Finding: Wang et al. (2023) showed chromatin remodeling activates cryptic sites
Evidence: [papers/Wang_2023_Cell.pdf]
---

---

Complex Logical Queries

MemU uses LLM-based search (not just embeddings), enabling complex logic:

# Find experiments where:
# - Chromatin was used
# - AND F1 > 0.9
# - AND tested on cancer samples
results = memory.query(
    "experiments with chromatin modality, F1 above 0.9, on cancer samples"
)

Why this matters: Semantic search would miss the logical constraints (F1 > 0.9, cancer-only).

---

🔬 Example: Novel Discovery Workflow

Full Workflow with Memory

from agentic_spliceai import MetaModel, NexusAgent, AgenticMemory

# Initialize
model = MetaModel()
agent = NexusAgent(memory=AgenticMemory())

# 1. Make predictions
predictions = model.predict(gene="BRCA1")

# 2. Find high-delta sites (potential novel isoforms)
novel_sites = [s for s in predictions if s.delta_score > 0.5]

# 3. Validate with memory context
for site in novel_sites:
    # Agent searches memory for similar cases
    similar = agent.memory.query(
        f"novel splice sites in {site.gene} with high chromatin signal"
    )

    # Agent validates with historical context
    validation = agent.validate(
        site,
        context=similar,  # ← Uses past discoveries!
        sources=["literature", "rnaseq", "databases"]
    )

    # 4. Store discovery if high confidence
    if validation.confidence > 0.8:
        agent.memory.store_discovery(
            site=site,
            evidence=validation.evidence,
            confidence=validation.confidence
        )

        print(f"✅ Novel site discovered: {site.location}")
        print(f"   Confidence: {validation.confidence}")
        print(f"   Evidence: {len(validation.evidence)} sources")
        print(f"   Similar cases: {len(similar)} in memory")

Output:

✅ Novel site discovered: chr17:43,094,582
   Confidence: 0.92
   Evidence: 4 sources (literature + RNA-seq + databases + chromatin)
   Similar cases: 3 in memory (BRCA1 exon 10, exon 12)

---

What Got Stored

File: memory/discoveries/brca1_novel_isoforms.md

# BRCA1 Novel Isoforms

## Discovery: BRCA1_exon11_cryptic_donor_20260220

**Location**: chr17:43,094,582 (GRCh38)  
**Type**: Cryptic donor (GT motif)  
**Discovery Date**: 2026-02-20

### Prediction Scores
- Base model: 0.15 (LOW - not in training data)
- Meta model: 0.88 (HIGH - strong multimodal evidence)
- **Delta: 0.73** (large gap → novel signal)

### Evidence
**Sequence**: AGGTAAGT (canonical GT donor, 0.90 strength)  
**Chromatin**: 0.82 (open in breast tissue)  
**Histone**: H3K36me3 = 0.76 (active transcription)  
**RNA-seq**: 24 junction reads in TCGA-BRCA  
**Literature**: Confirmed by Venkitaraman (2014, Nature)

### Validation
- **Confidence**: 0.92 (HIGH)
- **Agent**: nexus-v2
- **Status**: ✅ Validated
- **Clinical**: Likely pathogenic (familial breast cancer)

### Provenance
- [rnaseq/TCGA-BRCA/junction_counts.tsv]
- [papers/Venkitaraman_2014_Nature.pdf]
- [predictions/meta-v2/BRCA1_predictions.tsv]
- [igv_screenshots/BRCA1_exon11_junction.png]

---

🧪 Scientific Workflows

Workflow 1: Hypothesis Testing

# Propose hypothesis
agent.memory.store_hypothesis({
    "title": "High-delta sites predict novel isoforms",
    "definition": "Sites with delta > 0.5 are enriched for novel sites",
    "rationale": "Base model hasn't seen them, meta model has context",
    "proposed": "2026-01-30"
})

# Run experiments
for experiment in ["chr21", "brca1", "tp53"]:
    results = run_experiment(experiment)
    agent.memory.add_experiment_to_hypothesis(
        hypothesis_id="high_delta_predicts_novel",
        experiment=results
    )

# Query hypothesis status
status = agent.memory.query("status of high-delta hypothesis")
# → "Validated in 3/3 experiments. Precision = 0.85, Recall = 0.73"

---

Workflow 2: Error Pattern Analysis

# Analyze false positives
fps = model.get_false_positives(test_set="chr21")

# Store pattern
agent.memory.store_error_pattern({
    "type": "false_positive",
    "pattern": "High FP rate in repetitive regions",
    "frequency": 0.23,  # 23% of all FPs
    "cause": "Base model trained on canonical sites",
    "mitigation": "Add repeat masking + chromatin filter",
    "effectiveness": "78% FP reduction"
})

# Later experiments automatically benefit
agent.memory.query("how to reduce false positives?")
# → Returns stored mitigation strategies

---

Workflow 3: Cross-Study Validation

# Discovery in Study 1 (BRCA1)
agent.memory.store_discovery({
    "gene": "BRCA1",
    "site": "chr17:43,094,582",
    "evidence": ["TCGA-BRCA RNA-seq"],
    "confidence": 0.85
})

# Study 2 (TP53) - Agent checks for similar patterns
similar = agent.memory.query(
    "discoveries with cryptic donors and chromatin evidence"
)

# Agent: "Found similar pattern in BRCA1. Let me check TP53..."
# → Cross-validates across studies
# → Builds confidence through independent replication

---

📊 Benefits for Your Research

1. Time Savings

• No more searching through old log files
• No more re-reading papers you've already processed
• No more explaining context to tools each session

Estimate: Save 2-5 hours/week

---

2. Better Science

• Reproducibility: Full provenance for every claim
• Cumulative learning: Build on past experiments
• Cross-validation: Connect findings across studies
• Error avoidance: Remember what didn't work

---

3. Publication Ready

Memory files can directly become: - Methods sections (full experimental details) - Supplementary materials (all evidence chains) - Data availability statements (all artifacts linked)

---

4. Collaborative

• Share memory with lab members
• Build institutional knowledge
• Onboard new researchers faster

---

🚀 Getting Started

Installation

# Install Agentic-SpliceAI with memory support
pip install agentic-spliceai[memory]

# Initialize memory
agentic-spliceai memory init --dir ./memory

---

Basic Usage

from agentic_spliceai.memory import AgenticMemory

# Create memory
memory = AgenticMemory()

# Store something
memory.store(
    category="experiments",
    item={"name": "My first experiment", "result": "success"}
)

# Query it
results = memory.query("my first experiment")
print(results)  # Finds it!

---

CLI Commands

# Search memory
agentic-spliceai memory search \
  --query "chromatin modality effectiveness"

# Store experiment
agentic-spliceai memory store \
  --category experiments \
  --file experiment_results.json

# Export for paper
agentic-spliceai memory export \
  --categories discoveries,literature \
  --gene BRCA1 \
  --output supplementary_materials/

---

🎓 Best Practices

1. Store Early, Store Often

Don't wait until "final results"—store intermediate findings too.

Why: Negative results and failed experiments are valuable knowledge!

---

2. Always Link to Raw Data

Every memory item should link back to source files.

memory.store({
    "finding": "Chromatin improved F1 by 8%",
    "evidence": [
        "logs/experiment_123.log",  # ← Link to raw
        "results/metrics.json"
    ]
})

---

3. Use Descriptive Categories

Organize memory by scientific theme, not tool structure.

Good: memory/tissue_specific_splicing/
Bad: memory/model_outputs/

---

4. Review High-Stakes Memories

Validate discoveries before storing as "confirmed."

if validation.confidence > 0.9:
    memory.store(item, reviewed=True, reviewer="expert@uni.edu")

---

5. Query Before You Start

Check memory before running experiments.

# Before experiment
prior_work = memory.query("chromatin modality on chr21")
if prior_work:
    print("We tried this before! Here's what we learned:")
    print(prior_work)

---

📚 Advanced Topics

Memory Provenance Chains

Every claim traces back through full evidence chain:

Discovery: "BRCA1 exon 11 cryptic donor"
    ↓
Evidence: "24 RNA-seq junction reads"
    ↓
Experiment: "TCGA-BRCA analysis"
    ↓
Raw Data: "tcga_brca_junctions.bam"

Query any level:

memory.trace_provenance(discovery_id="BRCA1_exon11_cryptic")
# Returns full chain with file paths

---

Memory Versioning

Memory files are Markdown → version with Git:

git log memory/discoveries/brca1_novel_isoforms.md
# See how knowledge evolved over time

---

Collaborative Memory

Share memory across team:

# Clone lab memory
git clone lab-server:/memory ./shared_memory

# Use it
memory = AgenticMemory(root_dir="./shared_memory")

# Contribute discoveries
memory.store_discovery(...)
git commit -m "Added TP53 cryptic site validation"
git push

---

🔗 Learn More

• Framework: MemU (https://github.com/NevaMind-AI/memU)
• Use Cases: See docs/agency/MEMORY_PATTERNS.md
• API Reference: See docs/api/memory.md

---

🎉 Summary

Agentic memory transforms AI agents from tools into research collaborators that: - ✅ Remember what you've learned - ✅ Build cumulative knowledge - ✅ Provide full provenance - ✅ Get smarter over time

The result: Better science, faster discovery, stronger publications.

---

Try it today: pip install agentic-spliceai[memory]

Questions? Open an issue on GitHub or email research@agentic-spliceai.org

---

"The best research assistant is one that remembers everything you've taught it."