In this module

MF1.1 The Acquisition Problem — Smear, Volatility, and the Impossibility of a Perfect Capture

6 hours · Module 1 · Free
What you already know

From MF0 you know memory contains evidence disk does not — in-memory payloads, live credentials, cleartext buffers, kernel structures attackers manipulate. You've seen the tools (Volatility 3, MemProcFS, WinDbg), the workflow, the confidence tiers, and the legal frame. What MF0 assumed — and MF1 makes real — is that you can actually get the image in the first place. This sub covers why that's harder than it looks.

Operational Objective

Memory acquisition isn't a snapshot. It's a process that takes 30 seconds to 15 minutes while the system keeps running, keeps scheduling threads, keeps swapping pages, and keeps handling interrupts. The image you end up with describes a range of moments, not one moment — and until you internalise that, half the strange results you'll see in later modules look like tool bugs instead of the acquisition artifacts they actually are.

Most practitioners come into memory forensics assuming acquisition is the easy part. It looks like a one-liner: run WinPmem, get a .raw file, move to analysis. What they discover in production is that the one-liner ships a file whose structures don't fully agree with each other, whose process list may include a thread that exited during the capture, and whose timestamp footprint spans the several minutes the acquisition actually took. Every one of those artifacts is normal. None of them mean the tool failed. All of them matter for how you interpret what you find later.

This sub sets the frame for the rest of the module. The order of volatility explains why memory comes first and why you still lose some of it no matter how fast you move. Smear explains why the image is a range, not a point. The "no perfect acquisition" principle explains why your job is choosing the least-damaging method for this system in this moment, documenting the choice, and defending it later — not chasing a theoretical clean capture that no tool produces.

Deliverable: Working grasp of the order of volatility and where memory sits in it, the concept of acquisition smear and why it's unavoidable, the tradeoff space every acquisition method occupies (fidelity, footprint, feasibility), and the decision frame that determines which method is correct for a given target. You finish this sub understanding what you're actually doing when you capture memory — and why the rest of MF1 is mostly about managing tradeoffs rather than finding a flawless tool.

Estimated completion: 35 minutes
ORDER OF VOLATILITY — HOW FAST EVIDENCE DECAYS CPU REGISTERS, CACHES, ARP CACHE microseconds — effectively uncapturable RAM — running processes, network connections, credentials, encryption keys seconds to minutes — this is where memory forensics operates RUNNING-STATE FILES — pagefile, swap, tmpfs, open handles minutes to hours — survives until next boot DISK — filesystem, logs, registry hives, databases hours to years — the traditional forensic surface BACKUPS, ARCHIVES, OFFLINE STORAGE months to decades — durable but often stale RULE — capture top-down. Anything you skip on the volatile side is gone. The deeper you push toward RAM, the harder acquisition gets and the more evidence you collect.
Figure 1.1.1 — The order of volatility frames every acquisition decision. Memory acquisition sits in the orange band: fast-decaying, operationally demanding, and the highest-value target when attacker activity is live.

The order of volatility is a priority rule, not a theoretical ranking

This section establishes why memory is captured first when it matters. The rule is operational — it tells you what to do when you can't collect everything.

The order of volatility was formalised in RFC 3227 in 2002 and hasn't needed revising since. Some evidence types decay faster than others; when you can't collect everything, you collect the fastest-decaying first. Memory sits near the top. Disk sits near the bottom. An investigator responding to an active compromise who starts with disk imaging is doing it wrong — by the time the disk image completes, the memory that would have shown the running attacker process is gone.

Most SOC analysts who haven't done memory forensics read this as "memory first when possible." That's not what the rule says. The rule says memory first when memory matters at all, and memory matters whenever there's any chance the attacker is still present or recently was. Post-incident disk-only response — where the system was rebooted before the responder arrived, or where the business insisted on isolation before capture — throws away the memory evidence and makes the rest of the investigation harder. Investigations that succeed against modern attacker tradecraft are the ones where memory capture happened before anything else touched the system.

The tiers below RAM matter too. Pagefile and swap contain memory pages that were paged out — memory-adjacent, covered in MF1.5. Disk is the traditional forensic surface but also where attackers work hardest to leave no traces. Backups are durable but often snapshot a state from before the compromise. Each tier has different collection urgency and evidentiary value, and the order sharpens your thinking about where the evidence actually lives.

The rule also tells you when not to acquire memory. If the target has been offline for a week, memory is long gone and acquisition produces nothing. If the case is data exfiltration where the relevant evidence is network flows and file-access logs, memory is possibly useful but not time-critical. Non-capture is a valid decision when the evidence demands disappear.

Extended context — where RFC 3227 breaks down in 2026

RFC 3227 was written for a threat model that doesn't match modern attacker tradecraft. The RFC assumes post-incident evidence collection: an incident happened, the responder arrives later, they collect in volatility order. Modern incidents rarely work that way. Fileless malware, in-memory implants, and living-off-the-land execution mean memory is often the only place the attacker left traces. The order of volatility is still correct, but the stakes on the RAM tier have shifted — it's not just "fastest-decaying," it's "often the only source." That changes the operational calculus for whether to attempt acquisition at all, how much disruption to accept in exchange for the capture, and how aggressively to push back when management wants to reimage before capture completes.

The other drift is that RFC 3227 predates virtualisation. Hypervisor-based acquisition (MF1.4) wasn't an option when the RFC was written; it now offers a capture path with essentially zero footprint on the guest OS, which the RFC's in-OS acquisition assumption doesn't account for. The principle still holds; the methods have expanded.

Smear — the capture is a range of moments, not one moment

This section introduces the single most misunderstood property of memory images. Without understanding smear, every later analysis finding looks suspect for the wrong reasons.

When you run WinPmem against a live Windows system, the acquisition runs for tens of seconds to several minutes depending on RAM size and disk speed. During that time, the system keeps running. Processes get scheduled. Network packets arrive. Threads exit. The kernel updates structures. By the time the acquisition finishes writing the last page, the first page it wrote describes a system state that no longer exists — and the two pages disagree with each other in small but real ways. This is smear.

A concrete example. WinPmem walks physical memory from offset zero upward. It reaches the region containing the ActiveProcessLinks doubly-linked list at 15 seconds into the capture. At that moment, four processes are on the list. It writes those structures to the image. At 45 seconds, it reaches the region containing one of those processes' EPROCESS fields — but by then, the process has exited. The ActiveProcessLinks entry for it is stale in the image, but the EPROCESS fields are partially zeroed because the kernel started tearing the structure down. Volatility 3 reports the process with unusual field values. An analyst unfamiliar with smear sees this as weird, suspects a rootkit, and goes looking for what isn't there. The real explanation is prosaic: the acquisition straddled the process's exit.

Smear takes three forms. Structural smear — fields captured at different times don't fully agree. List smear — a linked list has a head captured early and tail captured late, with items reflecting different moments. Field smear — a single structure has some fields captured before a change and some after, producing impossible combinations (thread running with no CPU assigned, process marked terminated with live handles).

You cannot eliminate smear. The capture runs while the system runs. The only acquisition method that eliminates it entirely is freezing the system — suspending the VM at the hypervisor level (MF1.4), hibernating the OS so it writes a consistent snapshot to disk (MF1.5), or physically powering off the machine (which destroys everything you were trying to capture). Every live-system method produces some smear. Your job isn't to prevent it; it's to recognise it in later analysis and not waste investigation time on what it explains.

Three variables drive severity. Acquisition speed: faster capture means smaller smear window. System load: a busy system changes more state per second than an idle one. Memory size: larger RAM means longer capture means larger smear window for any given disk speed. MF1.7 covers smear detection and how to tell whether a particular capture is clean enough for the analysis you want to do.

There is no perfect acquisition — only tradeoffs you document

This section makes the key attitudinal shift the module needs you to make. Acquisition is a decision with tradeoffs, not a one-click operation.

Every acquisition method occupies a point in a three-dimensional tradeoff space. The axes are fidelity, footprint, and feasibility. No method maximises all three; every method sacrifices at least one.

Fidelity is how accurately the image represents the system's state at the time the capture started. Hypervisor suspension gives you near-perfect fidelity because the system is frozen during capture. Live OS acquisition with a software tool (WinPmem, LiME) gives you lower fidelity because smear happens during capture. Crash-dump acquisition gives you variable fidelity depending on the dump's completeness. The difference matters: a high-fidelity capture supports high-confidence findings; a low-fidelity capture caps your confidence tier.

Footprint is how much the acquisition changes the system during capture. A software capture tool loads into memory, which means the tool's own pages displace whatever pages would otherwise occupy that region — and those displaced pages might be exactly what you wanted. A kernel driver loading into the kernel pool shifts pool allocations. The tool's process shows up in the process list. Some acquisition artifacts are unavoidable (the tool has to run somewhere) but the footprint varies between methods, and the smaller the footprint, the less the capture interferes with the evidence.

Feasibility is whether the method is even available. Hypervisor suspension is the gold standard for fidelity and footprint, but it requires hypervisor access, which you don't have on a physical server or on a cloud VM where the hypervisor is the provider's. WinPmem requires administrator rights and the target running Windows. LiME requires kernel module loading, which matches the host kernel version, which you may not have pre-built. Feasibility constraints often make the highest-fidelity method impossible and force you down the list.

The mental model is a Pareto frontier: you're picking from a set of methods, each of which is dominated on some axes and dominant on others. Hypervisor-based wins on fidelity and footprint when it's feasible. Software tools win on feasibility when hypervisor access isn't available. Crash dumps win when the system already crashed and you're working with what's there. None is universally better. The acquisition decision is picking the method whose tradeoff profile best matches the case — and documenting why you picked it, because "why you picked it" is the first question opposing counsel asks in any investigation that reaches legal review.

The discipline this module is training: you don't pick the acquisition method by "best tool"; you pick by "best tradeoff for this situation." On an active attack where the adversary may notice the capture, footprint dominates. On a post-incident review where the attacker is long gone, fidelity dominates. On a cloud VM where the hypervisor is the provider's, feasibility dominates. Same course, same tools, different decisions depending on context. MF1.6 makes the decision explicit as a documented acquisition record; this sub establishes why the decision matters in the first place.

Guided Procedure — Classify three acquisition scenarios by tradeoff priority

Paper-based exercise. The goal is to internalise the fidelity-footprint-feasibility frame before MF1.2 starts covering specific tools. Read each scenario, decide which axis dominates, and name the method you'd pick. The "expected result" shows the canonical answer and reasoning.

Step 1 — Scenario A: active ransomware. A Northgate Engineering workstation has triggered Defender alerts consistent with ransomware pre-encryption staging. The attacker is assumed to be live on the host right now. The SOC lead asks for memory capture. You have local admin on the workstation and VMware vSphere access to the host hypervisor.
Expected reasoning: Footprint dominates. The attacker is live and may be watching for unusual processes, driver loads, or disk I/O that would signal incident response activity. Hypervisor suspension (MF1.4) has zero guest-visible footprint — the VM is frozen from outside. A software tool running in-guest loads a driver and writes several gigabytes to disk, both of which are visible to sophisticated malware.
If you picked WinPmem: You chose feasibility over footprint. WinPmem works and is faster to execute, but on an active ransomware case against an attacker who may be watching, the guest-visible activity can trigger the attacker's anti-IR response (fast-encrypt, C2 dump, self-delete). For this scenario, the hypervisor path is strictly better.
Step 2 — Scenario B: cloud VM post-incident review. A Northgate Engineering Azure VM showed anomalous outbound traffic three days ago. The SOC wants memory analysis to confirm or rule out in-memory implants. The VM is still running; Azure is the hypervisor; you have Azure RBAC contributor rights to the VM but no hypervisor access.
Expected reasoning: Feasibility dominates. You don't have hypervisor access — Azure doesn't expose `.vmem` extraction to tenants — so hypervisor suspension isn't available. The attacker is probably gone (three days post-incident), so footprint is less critical than it was in scenario A. You're back to in-guest software acquisition: WinPmem or equivalent. Document the method choice and the fact that hypervisor acquisition wasn't available.
If you picked hypervisor acquisition: You missed the feasibility constraint. Azure, AWS, and most public cloud providers don't give tenants raw hypervisor memory access. The highest-fidelity option is off the table; in-guest capture is what you've got.
Step 3 — Scenario C: lab baseline capture. Your Target-Win VM is freshly installed, no workload running, and you want a clean baseline memory image that MF2 onward will compare against. You have full VMware Workstation access to the hypervisor.
Expected reasoning: Fidelity dominates. There's no attacker, no time pressure, no footprint sensitivity. You want the cleanest possible reference image. Hypervisor suspension produces a `.vmem` file with no smear (the guest is frozen during capture). That's the baseline. If you wanted to verify the baseline using an in-guest method too, WinPmem against the suspended VM would let you compare — but the `.vmem` is the primary artifact.
If you picked WinPmem as the primary: You defaulted to the tool you're going to learn next instead of reasoning from the tradeoff frame. For a clean baseline with full hypervisor access and no constraints, `.vmem` is strictly better. Default to the best-available fidelity when nothing else dominates.
Step 4 — Extract the pattern. Across all three scenarios, the method didn't change based on what you knew about the tool. It changed based on the scenario's constraint profile. That's the decision frame MF1 is training.
Expected result: When you see an acquisition scenario in the rest of MF1 and across the paid modules, your first question is "which axis dominates here?" — not "what tool do I grab?" The tool falls out of the constraint analysis, not the other way around.
If you find yourself reaching for a tool first: Reread the "three-dimensional tradeoff space" section. The tool choice is an output of the constraint analysis, not an input to it.
Decision Point

The situation. A server in your environment triggered an alert an hour ago and the SOC has requested memory acquisition. The server is production — it handles the finance team's ERP connections during business hours, and it's 10:47 on a Tuesday. The server is a physical host (no hypervisor), running Windows Server 2022, with 128 GB of RAM. You have admin credentials. Finance needs the system up and usable by 13:00 for the afternoon invoice run. WinPmem against 128 GB will take approximately 12-15 minutes and will create a 128 GB file on the local disk (which the server has space for, but only just).

The choice. Proceed with WinPmem now, accepting the 12-15 minute capture during business hours. Delay until after 13:00, accepting that the attacker — if still live — has three more hours to move, exit, or cover tracks. Skip memory acquisition entirely and go straight to disk imaging, accepting that any fileless artifacts are lost.

The correct call. Proceed with WinPmem now, but announce it to finance and the SOC lead before starting. The order-of-volatility rule says you acquire while there's still something to acquire; three hours on a live attacker is enough time for the evidence to degrade substantially or disappear entirely. The 12-15 minute window is a real operational cost but not an existential one — finance's ERP connections slow during capture but don't break. Skipping acquisition entirely is the worst option: you've pre-committed to an investigation that can only see what's on disk, and if the attacker is fileless, the investigation fails from the start. The decision to capture now is documented in the acquisition record (MF1.6) along with the business-impact note and the SOC lead's acknowledgement; that documentation is the reason the capture is defensible later rather than second-guessed.

The operational lesson. Business constraint is an input to the acquisition decision, not a veto. Decisions that sacrifice evidence to business convenience without documentation are the ones opposing counsel exploits later — "the investigator chose not to capture memory on the grounds of a scheduled invoice run" is a finding pattern that doesn't survive adversarial review. Decisions that document the tradeoff honestly survive. Capture now, explain the business cost in the record, make the call the evidence requires.

Compliance Myth: "Memory acquisition tools produce reliable, consistent images the way disk imagers do"

The myth. Memory acquisition is the equivalent of disk imaging. You run the tool, it reads memory from start to finish, you get a bit-for-bit image of RAM the way dd produces a bit-for-bit image of a disk partition. The output is a clean, consistent file describing the system's state at the moment of capture.

The reality. Disk imaging and memory imaging share a file format but not much else. Disk imaging reads a storage medium whose contents aren't changing during the read (or at least, whose changes can be suppressed by write-blockers and mount-time discipline). Memory imaging reads a volatile substrate that keeps changing during the read itself, producing an image that describes a range of moments rather than one moment.

Every software-based memory acquisition introduces smear. The larger the RAM, the slower the disk, the busier the system, the worse the smear gets. A 128 GB capture on a busy server at 300 MB/s takes over seven minutes — during which thousands of processes may have started and exited, hundreds of millions of page faults may have been served, and the kernel's scheduling state will have cycled through millions of context switches. The final .raw file is not a snapshot; it's a stitched-together read of memory that was moving.

The operational implication is that you don't treat memory images with the same "bit-perfect" assumption you treat disk images with. Two memory captures of the same system taken minutes apart produce different .raw files with different SHA-256 hashes — not because the tool is broken, but because memory changed between the two runs. The hash still matters; it proves chain-of-custody integrity of the capture file, not perfect reproducibility of the source. MF1.6 covers this distinction in depth.

Practitioners who conflate the two imaging types write reports that overclaim. "The memory image is a forensically sound copy of the target's RAM at 14:22" sounds right but misrepresents what memory forensics tools produce. The accurate claim is "the memory image reflects the state of the target's RAM during the acquisition window of 14:22-14:29, with smear artifacts consistent with normal system load." The first version doesn't survive cross-examination; the second does.

Next

MF1.2 — Windows Acquisition with WinPmem. Now that you have the frame, MF1.2 walks through the most common Windows acquisition tool in detail: installation, execution, output format verification, and the common ways WinPmem fails. MF1.2 is where acquisition stops being conceptual and starts being operational — you will have captured a memory image of your Target-Win VM by the end of that sub.

Try it — Estimate smear window for your own analysis workstation

Setup. No tool install required for this exercise — just your analysis workstation and a calculator. You'll use published tool performance numbers and your workstation's hardware to estimate what a memory capture of the workstation itself would look like. The numbers are for reasoning about smear, not producing a capture.

Task. Calculate three numbers. First, the acquisition time: your workstation's RAM in gigabytes, divided by the sustained write speed in GB/s of the disk you'd write the capture to (modern NVMe SSDs sustain 1-3 GB/s for sequential writes; SATA SSDs sustain 400-500 MB/s; spinning disks sustain 100-150 MB/s). Second, an estimate of your workstation's context-switch rate during the capture — if you don't know, use 10,000 per second as a reasonable default for a moderately busy desktop. Third, the number of context switches that happen during the acquisition window — acquisition time in seconds multiplied by context-switch rate. That number is your smear budget.

Expected result. A 32 GB workstation with an NVMe SSD takes about 15-30 seconds to capture, during which 150,000 to 300,000 context switches occur. A 128 GB server with a SATA SSD takes 4-5 minutes, during which 2.4 to 3 million context switches occur. Both numbers are large. They make concrete what "smear" means operationally: the image you capture reflects hundreds of thousands to millions of scheduling transitions that happened while the capture was running.

If your result doesn't match. If your acquisition time estimate came out under 5 seconds, you either used a faster disk than your workstation actually has or forgot that acquisition writes the full RAM contents (not just used pages). If your smear budget came out under 10,000 context switches, you assumed a quieter system than realistic — modern desktops with browser tabs, IDE processes, and background services easily sustain 10,000+ switches per second even when they feel idle.

Checkpoint — before moving on

You should be able to do the following without referring back to this sub. If you can't, the sections to re-read are noted.

1. State in one sentence why memory is captured before disk in the order of volatility, and give one scenario where the rule would be wrong to apply rigidly. (§ The order of volatility is a priority rule)
2. Explain in one sentence what smear is, and give one concrete example of an analysis finding that would look suspicious to an analyst unfamiliar with smear but is actually a normal artifact of live-system capture. (§ Smear — the capture is a range of moments)
3. Name the three axes of the acquisition tradeoff space, and state which axis dominates in: (a) an active-attack scenario, (b) a cloud VM with no hypervisor access, (c) a clean lab baseline. (§ There is no perfect acquisition — only tradeoffs)

You've set up the lab and captured your first clean baselines.

MF0 built the three-VM lab and established the memory forensics landscape. MF1 taught acquisition with WinPmem and LiME, integrity verification, and chain of custody. From here, you execute attacks and investigate what they leave behind.

  • 8 attack modules (MF2–MF9) — process injection, credential theft, fileless malware, persistence, kernel drivers, Linux rootkits, timeline construction, and a multi-stage capstone
  • You run every attack yourself — from Kali against your target VMs, then capture memory and investigate your own attack's artifacts with Volatility 3
  • MF9 Capstone — multi-stage chain (initial access → privilege escalation → credential theft → persistence → data staging), three checkpoint captures, complete investigation report
  • The lab pack — PoC kernel driver and LKM rootkit source code, setup scripts, 21 exercises, 7 verification scripts, investigation report templates
  • Cross-platform coverage — Windows and Linux memory analysis in one course, with the timeline module integrating evidence from both
Unlock with Specialist — £25/mo See Full Syllabus

Cancel anytime